Merge PR #81: make the update checker visible — persistent banner + Help menu item (#78, #79)

Three conflicts, two of them semantic rather than textual:

- bcc.py QSS: this branch added the noticeBanner rules using the old
  module-level constants ({MUTED}, {ACCENT}); main had since moved the
  stylesheet onto palette slots ({p.muted}). Took main's form and
  translated the notice rules into it -- picking either side wholesale
  would have either dropped the banner styling or reintroduced globals
  that test_stylesheet_builder_has_no_hardcoded_colours now forbids.
- bcc.py methods: both sides appended to MainWindow (update-notice
  handlers vs theme handlers). Additive, kept both.
- tests/test_core.py: the usual EOF append. Kept both blocks.

_build_menu_bar auto-merged cleanly (View menu above, Help menu below);
verified both are present with their menu roles intact.

Verified: 272 test functions = 265 (main) + 7 (this branch), no
duplicates; 421 passed, ruff clean.

.github/workflows/ci.yml

@@ -68,3 +68,109 @@ jobs:
 
       - name: Run tests
         run: python -m pytest -v
+
+  # ── Catalog signature gate (#61) ─────────────────────────────────────────
+  #
+  # data/catalog.json is a list of command+args entries that BCC writes into
+  # the user's Claude config, which Claude then EXECUTES. The catalog is only
+  # trusted if it carries a valid Ed25519 signature from the maintainer key.
+  #
+  # The threat this gate exists for is NOT an outsider pushing to the repo —
+  # it is the maintainer merging a friendly-looking PR without really reading
+  # it. A contributor can change catalog.json but cannot produce a matching
+  # signature, so a blindly-merged PR lands here as a RED BUILD within a
+  # minute, instead of quietly riding into the next release.
+  #
+  # Public-key verification only. No secret is used or needed.
+  catalog-signature:
+    name: Catalog signature
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install dependencies
+        run: pip install cryptography
+
+      # 🔴 TRUST ANCHOR — issue #68 finding 4.
+      #
+      # This step used to do `import bcc_core as c` FROM THE CHECKED-OUT PR
+      # BRANCH and verify the catalog against c.CATALOG_PUBKEYS — i.e. it
+      # trusted the public key shipped in the very diff it was reviewing. A
+      # PR that changed data/catalog.json AND bcc_core.CATALOG_PUBKEYS (to
+      # an attacker key, with a matching signature produced by the attacker's
+      # matching private key) went green, because there was nothing outside
+      # the PR's own content to check the key against. The gate's whole
+      # point is catching a friendly-looking PR the maintainer merges
+      # without really reading it — and that hole made it a two-file diff.
+      #
+      # EXPECTED_CATALOG_PUBKEY_B64 below is hardcoded HERE, in the workflow
+      # file, independent of whatever bcc_core.py says on the PR branch. It
+      # is intentionally the only line in this step that matters for
+      # security review: changing it changes what this gate is willing to
+      # trust. THIS CONSTANT IS A TRUST ANCHOR. A PR that changes this line
+      # in the same diff as a catalog change is exactly the attack this gate
+      # exists to prevent — review a change to this line on its own,
+      # never bundled with a catalog update.
+      #
+      # NOTE for the next key rotation: update EXPECTED_CATALOG_PUBKEY_B64
+      # below to the new key's base64 form, as its own reviewed change.
+      - name: Verify data/catalog.json.sig
+        env:
+          EXPECTED_CATALOG_PUBKEY_B64: "082NOwVB7uURkvfyS3+knJ+40Fk6C9unsF47+2uPKo4="
+        run: |
+          python - <<'PY'
+          import base64, os, pathlib, sys
+          import bcc_core as c
+
+          expected_pubkey_b64 = os.environ["EXPECTED_CATALOG_PUBKEY_B64"]
+
+          raw = pathlib.Path("data/catalog.json").read_bytes()
+          sig_path = pathlib.Path("data/catalog.json.sig")
+
+          if not sig_path.exists():
+              sys.exit("FAIL: data/catalog.json.sig is missing. The catalog must be "
+                       "signed via the Catalog Console (#62) before it can land.")
+
+          if b"\x00" * 32 in c.CATALOG_PUBKEYS:
+              sys.exit("FAIL: CATALOG_PUBKEYS still holds the placeholder key.")
+
+          # Trust anchor check FIRST, before verifying anything against
+          # bcc_core.CATALOG_PUBKEYS: a PR is not allowed to bring its own
+          # key. CATALOG_PUBKEYS on the checked-out branch must be EXACTLY
+          # the key(s) this workflow file itself expects -- no more, no
+          # fewer, no substitutions.
+          actual_pubkeys_b64 = [base64.b64encode(k).decode() for k in c.CATALOG_PUBKEYS]
+          if actual_pubkeys_b64 != [expected_pubkey_b64]:
+              sys.exit(
+                  "FAIL: bcc_core.CATALOG_PUBKEYS on this branch does not match the "
+                  "trust anchor hardcoded in .github/workflows/ci.yml.\n"
+                  f"  expected: {[expected_pubkey_b64]}\n"
+                  f"  actual:   {actual_pubkeys_b64}\n"
+                  "\n"
+                  "This PR is changing (or has changed) the catalog signing key. That "
+                  "change must be reviewed on its own, separately from any catalog "
+                  "content change, and the workflow's EXPECTED_CATALOG_PUBKEY_B64 "
+                  "updated deliberately -- not accepted because it happened to match "
+                  "whatever bcc_core.py says on this branch."
+              )
+
+          if not c.verify_catalog_signature(raw, sig_path.read_bytes(), c.CATALOG_PUBKEYS):
+              sys.exit(
+                  "FAIL: data/catalog.json does NOT match its signature.\n"
+                  "\n"
+                  "The catalog changed without being re-signed. Either someone edited\n"
+                  "it directly (a PR you merged?), or a signing pass was forgotten.\n"
+                  "Re-review and re-sign with the Catalog Console — do not bypass this."
+              )
+
+          problems = c.validate_catalog(c.load_catalog(raw))
+          if problems:
+              sys.exit("FAIL: catalog failed validation:\n  " + "\n  ".join(problems))
+
+          print("OK: catalog signature verifies, the pubkey matches the CI trust anchor, "
+                "and the catalog validates clean.")
+          PY

.github/workflows/release.yml

@@ -95,6 +95,92 @@ jobs:
           name: ${{ matrix.artifact }}
           path: ${{ matrix.artifact }}
 
+  # ── Signing-key smoke test (workflow_dispatch only) ─────────────────────
+  #
+  # The Publish job is gated on a tag, so a manual run never exercises the
+  # signing step — which means a wrong/missing RELEASE_SIGNING_KEY secret
+  # would only be discovered at the worst possible moment: during a real
+  # release. This job signs a throwaway manifest with the secret and verifies
+  # the result against scripts/sign_checksums.RELEASE_PUBKEYS.
+  #
+  # IMPORTANT (issue #68 finding 5): this must verify against the RELEASE
+  # public key, never bcc_core.CATALOG_PUBKEYS. The catalog key is the
+  # offline, maintainer-held root of trust for what BCC executes; it must
+  # NEVER be compared against a value that lives in a CI secret, because
+  # that comparison is itself a way to smuggle a catalog-trusted key through
+  # CI review ("does this repo secret match the catalog key" is a question
+  # this workflow must never even ask). The release key is a SEPARATE
+  # keypair, generated via `catalog_console.py keygen --release`, that only
+  # ever signs release SHA256SUMS manifests -- a CI/secret compromise burns
+  # this key, not the catalog key.
+  #
+  # It proves the two halves of the RELEASE keypair actually match, without
+  # publishing anything. Run it from the Actions tab after setting or
+  # rotating the secret.
+  signing-smoke-test:
+    name: Signing key smoke test
+    runs-on: ubuntu-latest
+    if: github.event_name == 'workflow_dispatch'
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install dependencies
+        run: pip install cryptography
+
+      - name: Sign a throwaway manifest and verify against the RELEASE pubkey
+        env:
+          RELEASE_SIGNING_KEY: ${{ secrets.RELEASE_SIGNING_KEY }}
+        run: |
+          if [ -z "$RELEASE_SIGNING_KEY" ]; then
+            echo "FAIL: RELEASE_SIGNING_KEY secret is not set."
+            echo "Generate the RELEASE key (NOT the catalog key) with:"
+            echo "    python catalog_console.py keygen --release"
+            echo "then add its seed under Settings -> Actions -> Secrets, via:"
+            echo "    python catalog_console.py show-seed-b64 --release"
+            exit 1
+          fi
+          mkdir -p smoke && echo "smoke test payload" > smoke/hello.txt
+          python3 scripts/sign_checksums.py generate smoke --out smoke/SHA256SUMS
+          python3 scripts/sign_checksums.py sign --sums smoke/SHA256SUMS --out smoke/SHA256SUMS.sig
+          python - <<'PY'
+          import pathlib, sys
+          from scripts.sign_checksums import RELEASE_PUBKEYS, verify_checksums_against_any
+
+          # Deliberately does NOT import bcc_core / CATALOG_PUBKEYS at all --
+          # this smoke test must never be able to compare the CI secret
+          # against the catalog's root of trust (issue #68 finding 5). Only
+          # RELEASE_PUBKEYS (scripts/sign_checksums.py) is a legitimate
+          # target for a CI-resident key.
+          if not RELEASE_PUBKEYS:
+              sys.exit(
+                  "FAIL: scripts/sign_checksums.RELEASE_PUBKEYS is empty.\n"
+                  "\n"
+                  "Generate the release keypair with:\n"
+                  "    python catalog_console.py keygen --release\n"
+                  "then paste the printed public key into RELEASE_PUBKEYS in\n"
+                  "scripts/sign_checksums.py and commit that change."
+              )
+
+          sums = pathlib.Path("smoke/SHA256SUMS").read_text()
+          sig = pathlib.Path("smoke/SHA256SUMS.sig").read_bytes()
+
+          if not verify_checksums_against_any(RELEASE_PUBKEYS, sums, sig):
+              sys.exit(
+                  "FAIL: the signature produced by RELEASE_SIGNING_KEY does NOT verify\n"
+                  "against any key in scripts/sign_checksums.RELEASE_PUBKEYS.\n"
+                  "\n"
+                  "The secret and the shipped release public key are different keypairs.\n"
+                  "Downloaders would reject every signature this CI produces. Re-copy the\n"
+                  "seed from `catalog_console.py show-seed-b64 --release`, or update\n"
+                  "RELEASE_PUBKEYS with the matching public key."
+              )
+          print("OK: RELEASE_SIGNING_KEY matches a key in RELEASE_PUBKEYS.")
+          PY
+
   # ── Create GitHub Release with all three artifacts ──────────────────────
 
   release:
@@ -143,9 +229,13 @@ jobs:
       # checks. It does NOT remove Gatekeeper/SmartScreen warnings.
       #
       # The private key is a repo secret (RELEASE_SIGNING_KEY, base64 raw
-      # Ed25519 seed) generated via the Catalog Console (#62). If it's not
-      # set, we still publish the release — just without a .sig — rather
-      # than fail the release outright.
+      # Ed25519 seed) for the RELEASE key -- a SEPARATE keypair from the
+      # catalog key, generated via `python catalog_console.py keygen
+      # --release` (issue #68 finding 5; #62). This key is intentionally
+      # CI-resident and signs ONLY this checksum manifest; it is never
+      # trusted to sign data/catalog.json. If it's not set, we still
+      # publish the release — just without a .sig — rather than fail the
+      # release outright.
       - name: Check for signing key
         id: signing
         run: |
@@ -171,7 +261,7 @@ jobs:
       - name: Warn — release will be unsigned
         if: steps.signing.outputs.has_key != 'true'
         run: |
-          echo "::warning::RELEASE_SIGNING_KEY secret is not set — this release is being published WITHOUT a signed SHA256SUMS.sig. Add the secret (base64 raw Ed25519 seed, generated via the Catalog Console, #62) before the next tag."
+          echo "::warning::RELEASE_SIGNING_KEY secret is not set — this release is being published WITHOUT a signed SHA256SUMS.sig. Generate the RELEASE key (python catalog_console.py keygen --release) and add its seed (python catalog_console.py show-seed-b64 --release) as this secret before the next tag."
 
       - name: Create GitHub Release
         uses: softprops/action-gh-release@v2

README.md

@@ -36,11 +36,10 @@ are only suppressed by a paid OS-vendor certificate, which this project
 doesn't have. Verifying checksums is about detecting tampering in transit or
 on a mirror, not about vouching for the software.
 
-**Release signing public key** (Ed25519, base64, raw 32 bytes):
-
-```
-<PLACEHOLDER — AJ: paste the public key from the Catalog Console (#62) here>
-```
+This manifest is signed with BCC's **release key**, which is a different
+key from the one that signs the MCP server catalog — see
+[Signing keys](#signing-keys) below for why, and for the public key value
+to use with `--pubkey-b64` below.
 
 ### macOS / Linux
 
@@ -59,7 +58,7 @@ To also verify the manifest's signature (optional, requires Python +
 ```bash
 python3 scripts/sign_checksums.py verify \
   --sums SHA256SUMS --sig SHA256SUMS.sig \
-  --pubkey-b64 "<the public key above>"
+  --pubkey-b64 "<the release public key from Signing keys, below>"
 ```
 
 ### Windows (PowerShell)
@@ -78,6 +77,45 @@ release is missing the `.sig` file, the checksums themselves are still
 valid and safe to check against — the release workflow only skips signing,
 never checksum generation.
 
+## Signing keys
+
+BCC uses **two separate Ed25519 keypairs**, deliberately never the same
+key, because they protect different things and live in different places:
+
+| | Catalog key | Release key |
+|---|---|---|
+| Signs | `data/catalog.json` (the MCP server catalog every user's app trusts) | `SHA256SUMS` (the checksum manifest for release binaries) |
+| Verified by | `bcc_core.CATALOG_PUBKEYS` | `scripts/sign_checksums.RELEASE_PUBKEYS` |
+| Lives | Offline, passphrase-encrypted, maintainer's machine only (OS keychain or an encrypted file outside the repo — see the [Catalog Console](#files), issue #62) | A Gitea Actions repo secret, `RELEASE_SIGNING_KEY` — **intentionally CI-resident** |
+| Generated with | `python catalog_console.py keygen` | `python catalog_console.py keygen --release` |
+| Exported for CI with | *(never — there is no supported way to export this key)* | `python catalog_console.py show-seed-b64 --release` |
+
+**Why two keys:** the catalog key is the root of trust for what BCC
+actually *executes* on a user's machine — every `command`/`args` pair in
+the shipped catalog is only there because this key signed it. If that key
+and the release-checksum key were the same (as they briefly were — see
+[issue #68](../../issues/68)), then anything that can exfiltrate a Gitea
+Actions secret (a malicious workflow-file PR, a compromised runner, a leaky
+log) could sign a catalog every user's copy of BCC would trust, not just a
+checksum manifest. Splitting them means **a CI/secret compromise burns the
+release key, never the catalog key** — checksums for a future release could
+be forged, which is bad, but no attacker gains the ability to make BCC run
+arbitrary commands on installs that trust the catalog. That asymmetry is
+the entire point of having two keys instead of one.
+
+The catalog key is **never** meant to leave the maintainer's machine: it's
+generated, stored, unlocked, and used to sign entirely inside the Catalog
+Console (`catalog_console.py`), and `catalog_console.py show-seed-b64`
+refuses to run without `--release` specifically so the catalog seed can't
+be exported by habit or muscle memory.
+
+**Release signing public key** (Ed25519, base64, raw 32 bytes) — this is
+the RELEASE key, not the catalog key:
+
+```
+<PLACEHOLDER — AJ: paste the release public key from `catalog_console.py keygen --release` here>
+```
+
 ## Run from source
 
 ```bash
@@ -152,6 +190,7 @@ file is also listed, marked *legacy*, so you can copy them over.
 - `bcc.spec` — PyInstaller build spec (cross-platform).
 - `scripts/build_icons.py` — regenerates `icons/app.icns` and `icons/app.ico` from source PNGs.
 - `scripts/sign_checksums.py` — generates and Ed25519-signs the release `SHA256SUMS` manifest (see [Verifying your download](#verifying-your-download)).
+- `catalog_console.py` / `catalog_review.py` — **maintainer-only**, never shipped to users (excluded from `bcc.spec`; see `tests/test_catalog_console_packaging.py`). The Catalog Console: review + sign `data/catalog.json`, and generate/manage both signing keys (`keygen`, `keygen --release`) — see [Signing keys](#signing-keys).
 
 ## Building from source
 

bcc.py

@@ -10,6 +10,7 @@ Run:  python mcp_manager.py
 
 from __future__ import annotations
 
+import contextlib
 import sys
 import time
 from pathlib import Path
@@ -18,6 +19,7 @@ from typing import ClassVar
 from PySide6.QtCore import QRect, QSettings, QSize, Qt, QThread, QTimer, QUrl, Signal
 from PySide6.QtGui import (
     QAction,
+    QActionGroup,
     QColor,
     QCursor,
     QDesktopServices,
@@ -25,6 +27,7 @@ from PySide6.QtGui import (
     QIcon,
     QKeySequence,
     QPainter,
+    QPalette,
     QPixmap,
 )
 from PySide6.QtWidgets import (
@@ -65,85 +68,122 @@ import bcc_core as core
 # thread during drag-and-drop import, so skip anything larger than this.
 MAX_DROP_IMPORT_BYTES = 5 * 1024 * 1024  # 5 MB
 
-# --- One-line rebrand: change this to recolor the whole app --------------- #
-ACCENT = "#f97316"  # warm orange
-ACCENT_DIM = "#c2570b"
-BG = "#1b1d23"
-PANEL = "#23262e"
-PANEL_2 = "#2b2f39"
-TEXT = "#e7e9ee"
-MUTED = "#9aa0ad"
-BORDER = "#3a3f4b"
-GOOD = "#4ade80"
-BAD = "#f87171"
-WARN = "#fbbf24"
+# --- Theming (issue #75) -------------------------------------------------- #
+# The palette lives in bcc_core (testable without a Qt app); these module-level
+# names are rebound by `apply_palette()` whenever the theme changes.
+#
+# Why globals rather than passing a palette around: ~20 inline
+# `setStyleSheet(f"color: {MUTED}")` calls are scattered through this file, and
+# an f-string resolves its names when it runs, not when it's compiled. Rebinding
+# the globals means every one of those call sites picks up the new colour on its
+# next render, with no change to the call sites themselves.
+PALETTE = core.DARK_PALETTE
+ACCENT = ACCENT_DIM = BG = PANEL = PANEL_2 = TEXT = MUTED = BORDER = ""
+GOOD = BAD = WARN = REMOTE = ON_ACCENT = DISABLED_BG = MONO_BG = SEL_TEXT = ""
+STATUS_COLORS: dict[str, str] = {}
+HEALTH_COLORS: dict[str, str] = {}
 
-STATUS_COLORS = {"ok": GOOD, "missing": BAD, "warn": WARN, "remote": "#60a5fa", "unknown": WARN}
-STATUS_GLYPH = {"ok": "●", "missing": "●", "warn": "▲", "remote": "◆", "unknown": "○"}
+STATUS_GLYPH = {
+    "ok": "\u25cf",
+    "missing": "\u25cf",
+    "warn": "\u25b2",
+    "remote": "\u25c6",
+    "unknown": "\u25cb",
+}
 
 # Health dot (spawn-test outcome, see core.HealthStatus) shown per row in the
 # server tables' "Health" column -- distinct from the PATH-dependency Status
 # column above.
-HEALTH_COLORS = {"ok": GOOD, "failed": BAD, "untested": MUTED}
-HEALTH_GLYPH = {"ok": "●", "failed": "●", "untested": "○"}
+HEALTH_GLYPH = {"ok": "\u25cf", "failed": "\u25cf", "untested": "\u25cb"}
 
-STYLESHEET = f"""
+
+def build_stylesheet(p: core.Palette) -> str:
+    """Render the global QSS for a palette."""
+    return f"""
 /* No font-family here on purpose: Qt already uses the native system UI font
    on every platform (San Francisco / Segoe UI / desktop default). Naming
    web-CSS aliases like -apple-system forces a costly font-alias scan. */
-* {{ font-size: 13px; color: {TEXT}; }}
-QMainWindow, QDialog {{ background: {BG}; }}
+* {{ font-size: 13px; color: {p.text}; }}
+QMainWindow, QDialog {{ background: {p.bg}; }}
 QLabel#h1 {{ font-size: 15px; font-weight: 600; }}
-QLabel#muted {{ color: {MUTED}; }}
-QFrame#card {{ background: {PANEL}; border: 1px solid {BORDER}; border-radius: 10px; }}
+QLabel#muted {{ color: {p.muted}; }}
+QFrame#card {{ background: {p.panel}; border: 1px solid {p.border}; border-radius: 10px; }}
 QLineEdit, QPlainTextEdit, QComboBox {{
-    background: {PANEL_2}; border: 1px solid {BORDER}; border-radius: 7px;
-    padding: 6px 8px; selection-background-color: {ACCENT}; selection-color: #1a1205;
+    background: {p.panel_2}; border: 1px solid {p.border}; border-radius: 7px;
+    padding: 6px 8px; selection-background-color: {p.accent}; selection-color: {p.on_accent};
 }}
-QLineEdit:focus, QPlainTextEdit:focus, QComboBox:focus {{ border: 1px solid {ACCENT}; }}
+QLineEdit:focus, QPlainTextEdit:focus, QComboBox:focus {{ border: 1px solid {p.accent}; }}
 QComboBox::drop-down {{ border: none; width: 22px; }}
-QComboBox QAbstractItemView {{ background: {PANEL_2}; border: 1px solid {BORDER};
-    selection-background-color: {ACCENT}; outline: none; }}
-QPushButton {{ background: {PANEL_2}; border: 1px solid {BORDER}; border-radius: 7px;
+QComboBox QAbstractItemView {{ background: {p.panel_2}; border: 1px solid {p.border};
+    selection-background-color: {p.accent}; outline: none; }}
+QPushButton {{ background: {p.panel_2}; border: 1px solid {p.border}; border-radius: 7px;
     padding: 7px 13px; }}
-QPushButton:hover {{ border: 1px solid {ACCENT}; }}
-QPushButton:disabled {{ color: {MUTED}; background: {PANEL}; }}
-QPushButton#primary {{ background: {ACCENT}; border: 1px solid {ACCENT}; color: #1a1205; font-weight: 600; }}
-QPushButton#primary:hover {{ background: {ACCENT_DIM}; }}
-QPushButton#primary:disabled {{ background: {PANEL}; color: {MUTED}; border: 1px solid {BORDER}; }}
-QPushButton#danger:hover {{ border: 1px solid {BAD}; color: {BAD}; }}
-QTableWidget {{ background: {PANEL}; border: 1px solid {BORDER}; border-radius: 10px;
+QPushButton:hover {{ border: 1px solid {p.accent}; }}
+QPushButton:disabled {{ color: {p.muted}; background: {p.panel}; }}
+QPushButton#primary {{ background: {p.accent}; border: 1px solid {p.accent}; color: {p.on_accent}; font-weight: 600; }}
+QPushButton#primary:hover {{ background: {p.accent_dim}; }}
+QPushButton#primary:disabled {{ background: {p.panel}; color: {p.muted}; border: 1px solid {p.border}; }}
+QPushButton#danger:hover {{ border: 1px solid {p.bad}; color: {p.bad}; }}
+QTableWidget {{ background: {p.panel}; border: 1px solid {p.border}; border-radius: 10px;
     gridline-color: transparent; outline: none; }}
 QTableWidget::item {{ padding: 6px 8px; border: none; }}
-QTableWidget::item:selected {{ background: {ACCENT}; color: #1a1205; }}
+QTableWidget::item:selected {{ background: {p.accent}; color: {p.on_accent}; }}
 /* Inline cell editors: the global QLineEdit padding/radius clips the text
    inside a table row, so give editors a compact, flat style instead. */
 QTableWidget QLineEdit {{
-    background: {PANEL_2}; color: {TEXT}; border: 1px solid {ACCENT};
+    background: {p.panel_2}; color: {p.text}; border: 1px solid {p.accent};
     border-radius: 3px; padding: 0px 4px; margin: 0px;
-    selection-background-color: {ACCENT_DIM}; selection-color: #ffffff;
+    selection-background-color: {p.accent_dim}; selection-color: {p.selection_text};
 }}
-QHeaderView::section {{ background: {PANEL}; color: {MUTED}; border: none;
-    border-bottom: 1px solid {BORDER}; padding: 8px; font-weight: 600; }}
+QHeaderView::section {{ background: {p.panel}; color: {p.muted}; border: none;
+    border-bottom: 1px solid {p.border}; padding: 8px; font-weight: 600; }}
 QScrollBar:vertical {{ background: transparent; width: 10px; margin: 2px; }}
-QScrollBar::handle:vertical {{ background: {BORDER}; border-radius: 5px; min-height: 24px; }}
+QScrollBar::handle:vertical {{ background: {p.border}; border-radius: 5px; min-height: 24px; }}
 QScrollBar::add-line, QScrollBar::sub-line {{ height: 0; }}
-QLabel#statusbar {{ color: {MUTED}; padding: 4px 2px; }}
-QLabel#warnBanner {{ color: #1a1205; background: {WARN}; border-radius: 8px; padding: 8px 10px; font-weight: 600; }}
-QLabel#section {{ color: {MUTED}; font-weight: 600; font-size: 12px; padding: 2px 2px; }}
-QLabel#sectionDisabled {{ color: {MUTED}; font-weight: 600; font-size: 12px; padding: 2px 2px; }}
-QLabel#placeholder {{ color: {MUTED}; padding: 12px; background: {PANEL_2}; border: 1px dashed {BORDER}; border-radius: 8px; }}
-QTableWidget#disabledTable {{ background: #202229; }}
-QTableWidget#disabledTable::item:selected {{ background: {ACCENT}; color: #1a1205; }}
+QLabel#statusbar {{ color: {p.muted}; padding: 4px 2px; }}
+QLabel#warnBanner {{ color: {p.on_accent}; background: {p.warn}; border-radius: 8px; padding: 8px 10px; font-weight: 600; }}
+QFrame#noticeBanner {{ background: {p.panel_2}; border: 1px solid {p.accent}; border-radius: 8px; }}
+QLabel#noticeText {{ color: {p.text}; }}
+QPushButton#noticeClose {{ background: transparent; border: none; color: {p.muted}; font-size: 14px; padding: 2px; }}
+QPushButton#noticeClose:hover {{ color: {p.text}; }}
+QLabel#section {{ color: {p.muted}; font-weight: 600; font-size: 12px; padding: 2px 2px; }}
+QLabel#sectionDisabled {{ color: {p.muted}; font-weight: 600; font-size: 12px; padding: 2px 2px; }}
+QLabel#placeholder {{ color: {p.muted}; padding: 12px; background: {p.panel_2}; border: 1px dashed {p.border}; border-radius: 8px; }}
+QTableWidget#disabledTable {{ background: {p.disabled_bg}; }}
+QTableWidget#disabledTable::item:selected {{ background: {p.accent}; color: {p.on_accent}; }}
 QPlainTextEdit#diag {{ font-family: "Menlo", "Cascadia Code", "Consolas", "DejaVu Sans Mono", monospace;
-    font-size: 12px; background: #16181d; border: 1px solid {BORDER}; border-radius: 8px; }}
+    font-size: 12px; background: {p.mono_bg}; border: 1px solid {p.border}; border-radius: 8px; }}
 QFrame#diagCard {{ background: transparent; border: none; }}
 QSplitter::handle {{ background: transparent; }}
-QSplitter::handle:hover {{ background: {BORDER}; border-radius: 4px; }}
-QSplitter::handle:pressed {{ background: {ACCENT}; border-radius: 4px; }}
+QSplitter::handle:hover {{ background: {p.border}; border-radius: 4px; }}
+QSplitter::handle:pressed {{ background: {p.accent}; border-radius: 4px; }}
 """
 
 
+def apply_palette(p: core.Palette) -> str:
+    """Rebind the module-level colour names to `p` and return its stylesheet."""
+    global PALETTE, ACCENT, ACCENT_DIM, BG, PANEL, PANEL_2, TEXT, MUTED, BORDER
+    global GOOD, BAD, WARN, REMOTE, ON_ACCENT, DISABLED_BG, MONO_BG, SEL_TEXT
+    global STATUS_COLORS, HEALTH_COLORS
+    PALETTE = p
+    ACCENT, ACCENT_DIM = p.accent, p.accent_dim
+    BG, PANEL, PANEL_2 = p.bg, p.panel, p.panel_2
+    TEXT, MUTED, BORDER = p.text, p.muted, p.border
+    GOOD, BAD, WARN, REMOTE = p.good, p.bad, p.warn, p.remote
+    ON_ACCENT, DISABLED_BG, MONO_BG, SEL_TEXT = (
+        p.on_accent,
+        p.disabled_bg,
+        p.mono_bg,
+        p.selection_text,
+    )
+    STATUS_COLORS = {"ok": GOOD, "missing": BAD, "warn": WARN, "remote": REMOTE, "unknown": WARN}
+    HEALTH_COLORS = {"ok": GOOD, "failed": BAD, "untested": MUTED}
+    return build_stylesheet(p)
+
+
+STYLESHEET = apply_palette(core.DARK_PALETTE)
+
+
 # --------------------------------------------------------------------------- #
 # Background reachability tester (keeps the UI responsive during the request)
 # --------------------------------------------------------------------------- #
@@ -1493,6 +1533,54 @@ class AboutDialog(QDialog):
         QDesktopServices.openUrl(QUrl(self._release_url or core.RELEASES_URL))
 
 
+class NoticeBanner(QFrame):
+    """A persistent, dismissible notice with an optional action button.
+
+    The status bar is the wrong home for anything the user needs to act on --
+    21 call sites rewrite it, so a message posted there is gone by the next
+    click. That wiped the MSIX warning (#35) and then the update notice (#78).
+    This is the shared mechanism so it doesn't happen a third time.
+    """
+
+    def __init__(self, parent=None):
+        super().__init__(parent)
+        self.setObjectName("noticeBanner")
+        row = QHBoxLayout(self)
+        row.setContentsMargins(10, 8, 8, 8)
+        row.setSpacing(8)
+        self._label = QLabel("")
+        self._label.setObjectName("noticeText")
+        self._label.setWordWrap(True)
+        row.addWidget(self._label, 1)
+        self._action_btn = QPushButton("")
+        self._action_btn.setCursor(Qt.CursorShape.PointingHandCursor)
+        self._action_btn.hide()
+        row.addWidget(self._action_btn)
+        self._close_btn = QPushButton("\u2715")
+        self._close_btn.setObjectName("noticeClose")
+        self._close_btn.setCursor(Qt.CursorShape.PointingHandCursor)
+        self._close_btn.setFixedWidth(26)
+        self._close_btn.setToolTip("Dismiss")
+        self._close_btn.clicked.connect(self.hide)
+        row.addWidget(self._close_btn)
+        self.hide()
+
+    def show_notice(self, text: str, action_label: str = "", on_action=None):
+        self._label.setText(text)
+        self._label.setToolTip(text)
+        # Reconnect cleanly: a banner reused for a second notice would
+        # otherwise fire the previous notice's action too.
+        with contextlib.suppress(RuntimeError, TypeError):
+            self._action_btn.clicked.disconnect()
+        if action_label and on_action is not None:
+            self._action_btn.setText(action_label)
+            self._action_btn.clicked.connect(lambda _=False: on_action())
+            self._action_btn.show()
+        else:
+            self._action_btn.hide()
+        self.show()
+
+
 # --------------------------------------------------------------------------- #
 # Restart worker: core.restart_claude_desktop() blocks up to ~5 s on macOS
 # waiting for the old instance to exit, so it must run off the UI thread.
@@ -1550,6 +1638,11 @@ class MainWindow(QMainWindow):
         self.warn_banner.hide()
         root.addWidget(self.warn_banner)
 
+        # Update availability gets its own persistent banner rather than a
+        # status-line write, which the next UI action overwrites (#78).
+        self.update_banner = NoticeBanner(self)
+        root.addWidget(self.update_banner)
+
         # User-draggable divider between the server list and the editor.
         split = QSplitter(Qt.Orientation.Horizontal)
         split.setChildrenCollapsible(False)
@@ -1585,11 +1678,96 @@ class MainWindow(QMainWindow):
 
     # --- menu bar ---------------------------------------------------------- #
     def _build_menu_bar(self):
+        view_menu = self.menuBar().addMenu("&View")
+        theme_menu = view_menu.addMenu("Theme")
+        self._theme_group = QActionGroup(self)
+        self._theme_group.setExclusive(True)
+        current = stored_theme_setting()
+        for setting, label in (
+            (core.THEME_SYSTEM, "Match system"),
+            (core.THEME_LIGHT, "Light"),
+            (core.THEME_DARK, "Dark"),
+        ):
+            act = QAction(label, self, checkable=True)
+            act.setChecked(setting == current)
+            act.triggered.connect(lambda _checked=False, s=setting: self._set_theme(s))
+            self._theme_group.addAction(act)
+            theme_menu.addAction(act)
+
         help_menu = self.menuBar().addMenu("&Help")
+
+        # "Check for updates" used to exist only as a button inside the About
+        # dialog, which is not somewhere anyone looks for it (#79).
+        update_action = QAction("Check for updates…", self)
+        # Explicit role: macOS relocates actions it recognises by text, and
+        # some Qt versions treat "update" as application-menu material. Pin it
+        # so the item stays where the menu says it is on every platform.
+        update_action.setMenuRole(QAction.MenuRole.ApplicationSpecificRole)
+        update_action.triggered.connect(self.check_for_updates)
+        help_menu.addAction(update_action)
+        help_menu.addSeparator()
+
         about_action = QAction("About Better Claude Config…", self)
+        # Qt auto-assigns AboutRole to actions whose text starts with "About",
+        # which moves this into the application menu on macOS. That is the
+        # right home there -- state it explicitly rather than inheriting it by
+        # accident, since the behaviour is invisible from this call site.
+        about_action.setMenuRole(QAction.MenuRole.AboutRole)
         about_action.triggered.connect(self._show_about)
         help_menu.addAction(about_action)
 
+    def _show_update_notice(self, notice: dict):
+        """Surface an available update where it survives the next click."""
+        url = notice["url"]
+        self.update_banner.show_notice(
+            notice["text"],
+            action_label="Open releases page",
+            on_action=lambda: QDesktopServices.openUrl(QUrl(url)),
+        )
+
+    def check_for_updates(self):
+        """Menu-driven check. Unlike the startup check this is never throttled
+        and always reports back -- the user asked, so silence would read as a
+        broken button."""
+        self.status.setText("Checking for updates…")
+        self._menu_update_worker = UpdateCheckWorker()
+        self._menu_update_worker.done.connect(self._on_menu_update_checked)
+        self._menu_update_worker.start()
+
+    def _on_menu_update_checked(self, release: dict | None):
+        self._menu_update_worker = None
+        if release is None:
+            self.status.setText("Couldn't check for updates (offline?).")
+            return
+        QSettings("BCC", "BetterClaudeConfig").setValue("update/lastCheck", time.time())
+        notice = core.update_notice(core.__version__, release)
+        if notice:
+            self._show_update_notice(notice)
+            self.status.setText(f"Update available: {notice['version']}")
+        else:
+            self.update_banner.hide()
+            self.status.setText(f"You're up to date ({core.__version__}).")
+
+    def _set_theme(self, setting: str):
+        """Persist the theme choice and repaint the running window."""
+        QSettings("BCC", "BetterClaudeConfig").setValue("ui/theme", setting)
+        app = QApplication.instance()
+        if app is None:  # pragma: no cover - only in a headless test harness
+            return
+        app.setStyleSheet(theme_stylesheet_for(app, setting))
+        # The global stylesheet covers most of the UI, but the inline
+        # setStyleSheet calls (status dots, warning labels, update banner) only
+        # pick up the new palette when their widget next renders -- so re-render
+        # them now rather than leaving dark-on-light text behind.
+        self._repaint_themed_widgets()
+
+    def _repaint_themed_widgets(self):
+        """Re-run the inline-styled bits after a palette change."""
+        self.status.setStyleSheet(f"color: {MUTED};")
+        idx = self._current_index()
+        self._refresh_tables(select_index=idx if idx >= 0 else -1)
+        self._update_status(saved=False)
+
     def _show_about(self):
         AboutDialog(self).exec()
 
@@ -1610,10 +1788,9 @@ class MainWindow(QMainWindow):
         if release is None:
             return  # offline/failed check: don't advance lastCheck, allow retry
         QSettings("BCC", "BetterClaudeConfig").setValue("update/lastCheck", time.time())
-        if core.is_newer_version(core.__version__, release["version"]):
-            self.status.setText(
-                f"Update available: {release['version']}  ·  Help ▸ About to view it."
-            )
+        notice = core.update_notice(core.__version__, release)
+        if notice:
+            self._show_update_notice(notice)
 
     # --- layout persistence ---------------------------------------------- #
     def _restore_layout(self):
@@ -1939,9 +2116,21 @@ class MainWindow(QMainWindow):
                 return
             self.full_config = cfg
             repaired = True
+        # extract_servers tolerates malformed entries rather than raising (#72),
+        # but keep it inside the guard: a load failure must leave the previously
+        # loaded profile intact instead of half-swapping the window's state.
+        try:
+            servers = core.extract_servers(self.full_config)
+        except Exception as exc:  # pragma: no cover - defence in depth
+            QMessageBox.critical(
+                self,
+                "Could not read config",
+                f"{profile.path}\n\nThe server list couldn't be read: {exc}",
+            )
+            return
         self._loaded_stat = core.config_fingerprint(profile.path)
         self.current_profile = profile
-        self.servers = core.extract_servers(self.full_config)
+        self.servers = servers
         self.dirty = False
         self.restart_btn.hide()
         self._undo_stack.clear()
@@ -2264,7 +2453,7 @@ class MainWindow(QMainWindow):
         entry = self.servers[idx]
         old_name = entry.name
         entry.name = self.editor.current_name()
-        entry.data = self.editor.dump_data()
+        entry.set_data(self.editor.dump_data())
         # The server stays in its section (enable state unchanged), so update
         # its existing row in place rather than re-rendering.
         # An edit invalidates any cached "Test all" result -- the server that
@@ -2358,7 +2547,7 @@ class MainWindow(QMainWindow):
                 QMessageBox.StandardButton.Yes | QMessageBox.StandardButton.No,
             )
             if ans == QMessageBox.StandardButton.Yes:
-                self.servers[existing[name]].data = data
+                self.servers[existing[name]].set_data(data)
                 return False, True
             name = core.resolve_name_collision(name, {s.name for s in self.servers})
         self.servers.append(core.ServerEntry(name, data, True))
@@ -2478,6 +2667,11 @@ class MainWindow(QMainWindow):
                 except Exception as e:
                     QMessageBox.critical(self, "Reload failed", str(e))
                     return
+                # The reload above is the on-disk truth for everything the user
+                # didn't touch -- but it also wipes BCC-authored keys the user
+                # changed in this session (named sets), which apply_servers
+                # doesn't write. Carry them over before saving (#73).
+                contested = core.carry_owned_keys(self.full_config, fresh)
                 core.apply_servers(fresh, self.servers)
                 try:
                     backup = core.write_config(self.current_profile.path, fresh)
@@ -2490,8 +2684,13 @@ class MainWindow(QMainWindow):
                 self.dirty = False
                 self.save_btn.setEnabled(False)
                 bnote = f"  ·  backup: {backup.name}" if backup else "  ·  (new file)"
+                cnote = (
+                    f"  ·  kept your {', '.join(contested)} (the file on disk had a different copy)"
+                    if contested
+                    else ""
+                )
                 self.status.setText(
-                    f"Merged & saved {self.current_profile.path}{bnote}"
+                    f"Merged & saved {self.current_profile.path}{bnote}{cnote}"
                     f"  ·  Restart {self.current_profile.label} to apply."
                 )
                 self._offer_restart_button()
@@ -2649,6 +2848,33 @@ class MainWindow(QMainWindow):
         e.accept()
 
 
+def system_is_dark(app: QApplication) -> bool:
+    """Whether the desktop is currently using a dark appearance.
+
+    Read from the style's own window colour rather than per-platform APIs --
+    Qt has already resolved the OS appearance by the time it builds the
+    default palette, so this works the same on all three platforms.
+    """
+    try:
+        return app.palette().color(QPalette.ColorRole.Window).lightness() < 128
+    except Exception:  # pragma: no cover - defensive; never block startup on theming
+        return True
+
+
+def stored_theme_setting() -> str:
+    """The user's theme choice, defaulting to following the system."""
+    value = QSettings("BCC", "BetterClaudeConfig").value("ui/theme", core.THEME_SYSTEM)
+    return value if value in core.THEME_CHOICES else core.THEME_SYSTEM
+
+
+def theme_stylesheet_for(app: QApplication, setting: str | None = None) -> str:
+    """Resolve setting + OS appearance into a palette, apply it, return the QSS."""
+    if setting is None:
+        setting = stored_theme_setting()
+    theme = core.resolve_theme(setting, system_is_dark(app))
+    return apply_palette(core.palette_for(theme))
+
+
 def main():
     if sys.platform == "win32":
         # Without an explicit AppUserModelID, Windows taskbar groups the app
@@ -2667,7 +2893,7 @@ def main():
     icon = _app_icon()
     if not icon.isNull():
         app.setWindowIcon(icon)
-    app.setStyleSheet(STYLESHEET)
+    app.setStyleSheet(theme_stylesheet_for(app))
     win = MainWindow()
     win.show()
     sys.exit(app.exec())

bcc_core.py

@@ -13,7 +13,9 @@ in its original position.
 
 from __future__ import annotations
 
+import base64
 import contextlib
+import copy
 import difflib
 import functools
 import glob
@@ -48,6 +50,12 @@ DISABLED_KEY = "_disabledMcpServers"
 # parks the rest under DISABLED_KEY.
 SETS_KEY = "_bccServerSets"
 
+# Top-level keys BCC itself authors. They live in the client's config file, but
+# BCC is their owner, so on a stale-file merge the in-memory copy wins over the
+# on-disk one (see `carry_owned_keys`). Any future BCC-authored key belongs
+# here -- forgetting to add one is exactly how #73 happened.
+BCC_OWNED_KEYS = (SETS_KEY,)
+
 BACKUP_DIRNAME = ".bcc_backups"
 MAX_BACKUPS = 15
 
@@ -162,6 +170,39 @@ def fetch_latest_release(timeout: float = 4.0) -> dict | None:
     return {"version": tag, "url": payload.get("html_url") or RELEASES_URL}
 
 
+def update_notice(
+    current: str, release: dict | None, url_fallback: str = RELEASES_URL
+) -> dict | None:
+    """Decide whether to tell the user about a release, and what to say.
+
+    Returns {"version", "text", "url"} when `release` is newer than `current`,
+    or None when it isn't, when the check failed, or when the payload is
+    malformed. Kept here rather than in the GUI so the wording and the
+    should-we-notify decision are testable -- bcc.py can't be imported by the
+    test suite, which has no PySide6.
+
+    The text deliberately names no menu path. The old status-line notice read
+    "Help > About to view it", which is wrong on macOS: Qt relocates the About
+    action into the application menu (#79). A notice that carries its own
+    action can't drift out of sync with the platform.
+    """
+    if not isinstance(release, dict):
+        return None
+    version = release.get("version")
+    if not version or not isinstance(version, str):
+        return None
+    if not is_newer_version(current, version):
+        return None
+    # Tags carry a "v" prefix and __version__ doesn't; render both the same way
+    # so the notice doesn't read "Version v1.3.0 ... you're running 1.2.0".
+    shown = version.lstrip("vV")
+    return {
+        "version": version,
+        "text": f"Version {shown} is available. You're running {current.lstrip('vV')}.",
+        "url": release.get("url") or url_fallback,
+    }
+
+
 # --------------------------------------------------------------------------- #
 # Data model
 # --------------------------------------------------------------------------- #
@@ -177,16 +218,195 @@ class Profile:
         self.path = Path(self.path)
 
 
+class _NoRaw:
+    """Sentinel for ServerEntry.raw.
+
+    `None` can't do this job: `{"mcpServers": {"foo": null}}` is legal JSON and
+    a real malformed-entry case, so None has to mean "the config said null",
+    not "there was nothing here".
+    """
+
+    __slots__ = ()
+
+    def __repr__(self) -> str:  # keeps ServerEntry reprs readable in test output
+        return "<no raw>"
+
+
+NO_RAW = _NoRaw()
+
+
 @dataclass
 class ServerEntry:
+    """One server definition.
+
+    `data` is always a dict so every consumer can treat it as one. When the
+    config held something that wasn't a JSON object for this server (a string,
+    a number, a list -- all legal JSON, all wrong here), `data` is empty and
+    the original value is preserved verbatim in `raw` so Save round-trips it
+    instead of silently deleting the user's line. `lint_servers` surfaces it.
+    `raw` defaults to the NO_RAW sentinel rather than None, because a config
+    value of literal `null` is itself a malformed entry worth preserving.
+
+    Assigning `data` means the user replaced the definition through the editor,
+    which retires `raw` -- use `set_data` so that can't be forgotten.
+    """
+
     name: str
     data: dict
     enabled: bool = True
+    raw: object = NO_RAW
 
     @property
     def kind(self) -> str:
         return "remote" if "url" in self.data and "command" not in self.data else "stdio"
 
+    @property
+    def malformed(self) -> bool:
+        """True when the config value for this server wasn't a JSON object."""
+        return self.raw is not NO_RAW
+
+    def set_data(self, data: dict) -> None:
+        """Replace the definition from the editor, clearing any malformed original."""
+        self.data = data
+        self.raw = NO_RAW
+
+    def config_value(self):
+        """What to write back to the config: the edited dict, or the untouched
+        malformed original when the user never edited it."""
+        return self.data if self.raw is NO_RAW else self.raw
+
+
+# --------------------------------------------------------------------------- #
+# Theming (issue #75)
+# --------------------------------------------------------------------------- #
+THEME_SYSTEM = "system"
+THEME_LIGHT = "light"
+THEME_DARK = "dark"
+THEME_CHOICES = (THEME_SYSTEM, THEME_LIGHT, THEME_DARK)
+
+
+@dataclass(frozen=True)
+class Palette:
+    """Every colour the UI draws with.
+
+    Deliberately exhaustive: the stylesheet used to inline a handful of
+    near-black literals (`#1a1205` for text on the accent, `#202229` for the
+    disabled table, `#16181d` for the diagnostics pane), which is fine while
+    there's one theme and invisible breakage the moment there are two. Each
+    gets a slot here so a light palette can't silently inherit a dark value.
+    """
+
+    name: str
+    accent: str
+    accent_dim: str
+    bg: str
+    panel: str
+    panel_2: str
+    text: str
+    muted: str
+    border: str
+    good: str
+    bad: str
+    warn: str
+    remote: str
+    on_accent: str  # text drawn on top of an accent fill
+    disabled_bg: str  # the parked-servers table
+    mono_bg: str  # diagnostics / log panes
+    selection_text: str
+
+
+# The shipping theme through v1.3.0. These values are carried over verbatim --
+# adding a light theme must not restyle the dark one.
+DARK_PALETTE = Palette(
+    name="dark",
+    accent="#f97316",
+    accent_dim="#c2570b",
+    bg="#1b1d23",
+    panel="#23262e",
+    panel_2="#2b2f39",
+    text="#e7e9ee",
+    muted="#9aa0ad",
+    border="#3a3f4b",
+    good="#4ade80",
+    bad="#f87171",
+    warn="#fbbf24",
+    remote="#60a5fa",
+    on_accent="#1a1205",
+    disabled_bg="#202229",
+    mono_bg="#16181d",
+    selection_text="#ffffff",
+)
+
+# The semantic colours are NOT the dark ones lightened. #4ade80 / #fbbf24 sit
+# around 1.7:1 against white -- illegible. These are darkened to clear 4.5:1,
+# which `test_light_palette_meets_contrast` enforces so nobody "tidies" them
+# back toward the dark hues later.
+LIGHT_PALETTE = Palette(
+    name="light",
+    accent="#c2410c",
+    accent_dim="#9a3412",
+    bg="#f6f7f9",
+    panel="#ffffff",
+    panel_2="#eef0f4",
+    text="#1b1d23",
+    muted="#5c6270",
+    border="#d3d7de",
+    good="#15803d",
+    bad="#b91c1c",
+    warn="#a16207",
+    remote="#1d4ed8",
+    on_accent="#ffffff",
+    disabled_bg="#e9ebef",
+    mono_bg="#f0f2f5",
+    selection_text="#ffffff",
+)
+
+PALETTES = {DARK_PALETTE.name: DARK_PALETTE, LIGHT_PALETTE.name: LIGHT_PALETTE}
+
+
+def resolve_theme(setting: str, system_is_dark: bool) -> str:
+    """Map a stored theme setting + the OS appearance onto a concrete palette name.
+
+    Anything unrecognised (a hand-edited QSettings value, a setting written by
+    a future version) falls back to following the system rather than to a
+    fixed theme -- the user's desktop is the better guess.
+    """
+    if setting == THEME_DARK:
+        return THEME_DARK
+    if setting == THEME_LIGHT:
+        return THEME_LIGHT
+    return THEME_DARK if system_is_dark else THEME_LIGHT
+
+
+def palette_for(theme: str) -> Palette:
+    """Concrete palette by name; unknown names fall back to dark (the historical look)."""
+    return PALETTES.get(theme, DARK_PALETTE)
+
+
+def _hex_to_rgb(value: str) -> tuple[int, int, int]:
+    v = value.lstrip("#")
+    if len(v) == 3:
+        v = "".join(ch * 2 for ch in v)
+    return int(v[0:2], 16), int(v[2:4], 16), int(v[4:6], 16)
+
+
+def relative_luminance(color: str) -> float:
+    """WCAG relative luminance for a #rrggbb colour."""
+
+    def chan(c: int) -> float:
+        srgb = c / 255.0
+        return srgb / 12.92 if srgb <= 0.04045 else ((srgb + 0.055) / 1.055) ** 2.4
+
+    r, g, b = (chan(c) for c in _hex_to_rgb(color))
+    return 0.2126 * r + 0.7152 * g + 0.0722 * b
+
+
+def contrast_ratio(fg: str, bg: str) -> float:
+    """WCAG contrast ratio between two #rrggbb colours (1.0 to 21.0)."""
+    a, b = relative_luminance(fg), relative_luminance(bg)
+    lighter, darker = max(a, b), min(a, b)
+    return (lighter + 0.05) / (darker + 0.05)
+
 
 # --------------------------------------------------------------------------- #
 # Discovery
@@ -451,13 +671,30 @@ def repair_config_file(path: str | os.PathLike) -> tuple[dict, list[str], str]:
     return obj, notes, pretty
 
 
+def _server_entry(name: str, data, enabled: bool) -> ServerEntry:
+    """Build a ServerEntry, tolerating a value that isn't a JSON object.
+
+    A hand-edited config can legally hold `{"mcpServers": {"foo": "oops"}}` --
+    valid JSON, wrong shape. Calling dict() on that raises, which used to take
+    the whole load down before the linter ever got a look at it (#72). Keep the
+    original instead and let the linter report it.
+    """
+    if isinstance(data, dict):
+        return ServerEntry(name=name, data=dict(data), enabled=enabled)
+    return ServerEntry(name=name, data={}, enabled=enabled, raw=data)
+
+
 def extract_servers(cfg: dict) -> list[ServerEntry]:
-    """Pull enabled (`mcpServers`) and disabled (`_disabledMcpServers`) servers."""
+    """Pull enabled (`mcpServers`) and disabled (`_disabledMcpServers`) servers.
+
+    Never raises on a structurally-odd config -- malformed entries come back as
+    empty-data entries carrying their original value (see `_server_entry`).
+    """
     out: list[ServerEntry] = []
     for name, data in (cfg.get("mcpServers") or {}).items():
-        out.append(ServerEntry(name=name, data=dict(data), enabled=True))
+        out.append(_server_entry(name, data, True))
     for name, data in (cfg.get(DISABLED_KEY) or {}).items():
-        out.append(ServerEntry(name=name, data=dict(data), enabled=False))
+        out.append(_server_entry(name, data, False))
     return out
 
 
@@ -549,8 +786,8 @@ def apply_servers(cfg: dict, servers: list[ServerEntry]) -> dict:
     Write the server list back into `cfg` in place, preserving every other key
     and the position of `mcpServers`. Returns the same dict for convenience.
     """
-    enabled = {s.name: s.data for s in servers if s.enabled}
-    disabled = {s.name: s.data for s in servers if not s.enabled}
+    enabled = {s.name: s.config_value() for s in servers if s.enabled}
+    disabled = {s.name: s.config_value() for s in servers if not s.enabled}
 
     cfg["mcpServers"] = enabled  # replaces value if key existed; appends otherwise
     if disabled:
@@ -563,6 +800,34 @@ def apply_servers(cfg: dict, servers: list[ServerEntry]) -> dict:
 # --------------------------------------------------------------------------- #
 # Write (atomic, with rotating backups)
 # --------------------------------------------------------------------------- #
+def carry_owned_keys(local_cfg: dict, fresh_cfg: dict) -> list[str]:
+    """Carry BCC-authored top-level keys from `local_cfg` onto `fresh_cfg`.
+
+    Used by the stale-file "Merge & save" path, which reloads the file from
+    disk and re-applies the user's server edits. That reload used to drop
+    anything BCC owns but `apply_servers` doesn't write -- named server sets
+    vanished without a word (#73). BCC owns these keys, so the in-memory copy
+    wins; mutates `fresh_cfg` in place.
+
+    Returns the keys where the on-disk copy differed and was overwritten, so
+    the caller can tell the user something was actually contested rather than
+    merely carried across.
+
+    Deliberately one-directional: a key absent locally is left alone on disk.
+    We can't tell "user deleted their last set" from "user never had sets and
+    another machine just added some", and silently deleting someone else's
+    data is the worse of the two failures.
+    """
+    conflicts: list[str] = []
+    for key in BCC_OWNED_KEYS:
+        if key not in local_cfg:
+            continue
+        if key in fresh_cfg and fresh_cfg[key] != local_cfg[key]:
+            conflicts.append(key)
+        fresh_cfg[key] = copy.deepcopy(local_cfg[key])
+    return conflicts
+
+
 def _make_backup(path: Path) -> Path:
     bdir = path.parent / BACKUP_DIRNAME
     bdir.mkdir(exist_ok=True)
@@ -1379,9 +1644,21 @@ def lint_server(name: str, data: dict) -> list[str]:
 
 
 def lint_servers(servers: list[ServerEntry]) -> list[str]:
-    """Concatenate lint_server warnings across every entry, in order."""
+    """Concatenate lint_server warnings across every entry, in order.
+
+    Entries whose config value wasn't a JSON object at all are reported here
+    rather than in lint_server, which takes an already-dict `data` (#72).
+    """
     out: list[str] = []
     for s in servers:
+        if s.malformed:
+            nm = s.name.strip() or "(unnamed)"
+            out.append(
+                f"'{nm}': server definition is not an object "
+                f"(found {type(s.raw).__name__}) -- it is preserved as-is; "
+                f"edit it to replace it with a proper definition"
+            )
+            continue
         out.extend(lint_server(s.name, s.data))
     return out
 
@@ -2163,12 +2440,36 @@ def restart_claude_desktop() -> RestartResult:
 # is rejected by validate_catalog() regardless of how plausible it looks.
 CATALOG_ALLOWED_COMMANDS = frozenset({"npx", "uvx", "docker", "node", "python", "python3"})
 
+# Env var keys a catalog entry's config.env must never set. Every one of
+# these is a loader/interpreter override that lets a value walk straight
+# past CATALOG_ALLOWED_COMMANDS and the -e/--eval/-c deny-rule below: e.g.
+# NODE_OPTIONS="--require /tmp/x.js" turns an allowlisted `npx` entry into
+# arbitrary code execution without ever touching config.args, which is the
+# only field the allowlist/deny-rules/ASCII/secret checks used to cover.
+# Matched case-insensitively -- env keys are case-sensitive on POSIX, but a
+# `node_options` lookalike is exactly the kind of thing this must catch.
+CATALOG_DENIED_ENV_KEYS = frozenset(
+    {
+        "NODE_OPTIONS",
+        "PYTHONSTARTUP",
+        "PYTHONPATH",
+        "PYTHONHOME",
+        "LD_PRELOAD",
+        "LD_LIBRARY_PATH",
+        "DYLD_INSERT_LIBRARIES",
+        "DYLD_LIBRARY_PATH",
+        "BROWSER",
+        "PATH",
+        "NODE_REPL_EXTERNAL_MODULE",
+    }
+)
+
 # Ed25519 public keys allowed to sign a catalog, raw 32-byte form. A LIST
 # (not a single key) so keys can be rotated without bricking installs that
 # still trust an older key: verify_catalog_signature() accepts a match
 # against ANY key in this list.
 CATALOG_PUBKEYS: list[bytes] = [
-    b"\x00" * 32,  # TODO: real key from Catalog Console (#62)
+    base64.b64decode("082NOwVB7uURkvfyS3+knJ+40Fk6C9unsF47+2uPKo4="),
 ]
 
 # Domain-separation prefix for the signed message. The signature covers
@@ -2188,6 +2489,36 @@ _CATALOG_SECRET_ARG_RE = re.compile(r"(?i)--api[-_]?key=|--token=|--password=")
 # <PLACEHOLDER>-style tokens the GUI must have the user fill in before Save.
 _PLACEHOLDER_RE = re.compile(r"<[^<>\s]+>")
 
+# A catalog entry's id becomes an mcpServers JSON key AND is interpolated
+# into Qt.AutoText widgets (status bar, QMessageBox) -- an id like
+# "<b>Verified</b>" renders as markup there. Not RCE, but UI spoofing, so
+# ids are constrained to a plain lowercase slug.
+_CATALOG_ID_RE = re.compile(r"^[a-z0-9][a-z0-9._-]{0,63}$")
+
+# Docker flags that consume the next arg as a value (so that value must not
+# be mistaken for the image reference when locating it in config.args).
+_DOCKER_VALUE_FLAGS = frozenset(
+    {
+        "-e",
+        "--env",
+        "-v",
+        "--volume",
+        "-p",
+        "--publish",
+        "--name",
+        "-w",
+        "--workdir",
+        "-u",
+        "--user",
+        "--entrypoint",
+        "--network",
+        "--platform",
+        "--add-host",
+        "-l",
+        "--label",
+    }
+)
+
 # How many versions a single accepted catalog jump may leap in one go. Bounds
 # a "freeze" attack: a compromised/leaked signing key claiming an absurd
 # future version would otherwise permanently outrank every legitimate
@@ -2262,6 +2593,121 @@ def _docker_arg_violations(tag: str, args: list[str]) -> list[str]:
     return problems
 
 
+def _catalog_package_spec_version(spec: str) -> str | None:
+    """
+    Extract the version pin from an npm-style package spec, or None if the
+    spec carries no pin.
+
+    Handles unscoped "name@version" and scoped "@scope/name@version" --
+    scoped names have a leading "@" that is NOT the version separator, so a
+    naive split on the first/only "@" misparses "@scope/pkg" (no version)
+    as pinned to "scope/pkg". Splitting from the right side instead is safe
+    for both forms because a package name may contain "@" only as the
+    scope's leading character.
+    """
+    if spec.startswith("@"):
+        rest = spec[1:]
+        if "@" not in rest:
+            return None
+        _, _, version = rest.rpartition("@")
+        return version or None
+    if "@" not in spec:
+        return None
+    _, _, version = spec.rpartition("@")
+    return version or None
+
+
+def _catalog_package_spec_pinned(spec: str) -> bool:
+    """
+    True if `spec` carries an exact version pin. Covers npm's "name@version"
+    / "@scope/name@version" and uv's documented PyPI pin forms
+    "name@version" and "name==version".
+    """
+    if "==" in spec:
+        _, _, version = spec.partition("==")
+        return bool(version)
+    return bool(_catalog_package_spec_version(spec))
+
+
+def _first_catalog_package_spec(args: list[str]) -> str | None:
+    """
+    The first arg that could plausibly BE a package spec: skip flags
+    (leading "-") and <PLACEHOLDER> tokens (which can't be validated and
+    are filled in by the user later, never shipped by the catalog as the
+    package name itself). Everything after the first hit is ignored --
+    trailing flags, paths, and placeholders are not package specs.
+    """
+    for a in args:
+        if a.startswith("-"):
+            continue
+        if _PLACEHOLDER_RE.fullmatch(a):
+            continue
+        return a
+    return None
+
+
+def _catalog_pin_violations(tag: str, command: str, args: list[str]) -> list[str]:
+    """
+    Version-pinning enforcement (finding #3): a catalog PR can otherwise
+    ship `npx -y @scope/pkg` or `docker run img:latest` and the *next*
+    resolve of that package/image is whatever the registry serves that day
+    -- outside review, outside the signature's meaning. This is the only
+    place that enforces pinning at runtime; catalog_review.py's
+    risk_unpinned_package() is a maintainer-facing hint, not a gate.
+    """
+    if command in ("npx", "uvx"):
+        spec = _first_catalog_package_spec(args)
+        if spec is None:
+            return [f"{tag}: config.args must include a package spec to pin (e.g. name@1.2.3)."]
+        if not _catalog_package_spec_pinned(spec):
+            return [
+                f"{tag}: config.args package {spec!r} is not version-pinned; use "
+                "name@version, @scope/name@version, or name==version."
+            ]
+        return []
+
+    if command == "docker":
+        image = _docker_image_ref(args)
+        if image is None:
+            return [f"{tag}: config.args docker command has no image reference to pin."]
+        _, sep, image_tag = image.rpartition(":")
+        if not sep or "/" in image_tag:
+            return [
+                f"{tag}: config.args docker image {image!r} has no explicit tag; "
+                "pin an exact version (not 'latest', not untagged)."
+            ]
+        if image_tag == "latest":
+            return [
+                f"{tag}: config.args docker image {image!r} uses the 'latest' tag, "
+                "which is not allowed; pin an exact version."
+            ]
+        return []
+
+    return []
+
+
+def _docker_image_ref(args: list[str]) -> str | None:
+    """
+    Locate the image reference in a `docker run ...` args list: skip the
+    "run" subcommand and any flags, including ones that consume the next
+    token as a value (-e, -v, --name, ...) so that value isn't mistaken for
+    the image. The first remaining positional token is the image.
+    """
+    i = 0
+    if i < len(args) and args[i] == "run":
+        i += 1
+    while i < len(args):
+        a = args[i]
+        if a.startswith("-"):
+            if a in _DOCKER_VALUE_FLAGS and "=" not in a:
+                i += 2
+            else:
+                i += 1
+            continue
+        return a
+    return None
+
+
 def _validate_catalog_config(tag: str, config) -> list[str]:
     """Validate the `config` block of a basic-tier catalog entry."""
     if not isinstance(config, dict):
@@ -2282,10 +2728,11 @@ def _validate_catalog_config(tag: str, config) -> list[str]:
             f"({', '.join(sorted(CATALOG_ALLOWED_COMMANDS))})."
         )
 
-    args = config.get("args")
-    if not isinstance(args, list) or not all(isinstance(a, str) for a in args):
+    raw_args = config.get("args")
+    args_ok = isinstance(raw_args, list) and all(isinstance(a, str) for a in raw_args)
+    args = raw_args if args_ok else []
+    if not args_ok:
         problems.append(f"{tag}: config.args must be a list of strings.")
-        args = []
 
     for a in args:
         if not a.isascii():
@@ -2304,11 +2751,55 @@ def _validate_catalog_config(tag: str, config) -> list[str]:
     if command == "docker":
         problems.extend(_docker_arg_violations(tag, args))
 
+    # Version pinning (finding #3) -- only meaningful once command/args are
+    # actually well-formed; a malformed args list already got its own
+    # problem above and has nothing left to pin-check.
+    if args_ok and command in ("npx", "uvx", "docker"):
+        problems.extend(_catalog_pin_violations(tag, command, args))
+
     env = config.get("env")
-    if env is not None and (
-        not isinstance(env, dict) or any(not isinstance(v, str) for v in env.values())
-    ):
-        problems.append(f"{tag}: config.env must be an object of string values.")
+    if env is not None:
+        env_ok = isinstance(env, dict) and all(
+            isinstance(k, str) and isinstance(v, str) for k, v in env.items()
+        )
+        if not env_ok:
+            problems.append(f"{tag}: config.env must be an object of string values.")
+        else:
+            # config.env (finding #2): unlike args, env was previously
+            # type-checked ONLY -- no allowlist, no deny-rule, no ASCII
+            # check, no secret check. That made it the single easiest way
+            # to smuggle a payload past every other guard in this
+            # function: an allowlisted `command: npx` plus
+            # NODE_OPTIONS=--require /tmp/x.js in env walks straight past
+            # the command allowlist AND the -e/--eval/-c deny-rule above,
+            # because neither of those ever looks at env.
+            for key, value in env.items():
+                if not key.isascii():
+                    problems.append(
+                        f"{tag}: config.env key {key!r} must be ASCII "
+                        "(non-ASCII code points rejected)."
+                    )
+                if key.upper() in CATALOG_DENIED_ENV_KEYS:
+                    problems.append(
+                        f"{tag}: config.env key {key!r} is on the catalog deny-list "
+                        "(interpreter/loader override) and is not allowed."
+                    )
+                if not value.isascii():
+                    problems.append(
+                        f"{tag}: config.env value for {key!r} must be ASCII "
+                        "(non-ASCII code points rejected)."
+                    )
+                if _is_secret_value(value):
+                    problems.append(
+                        f"{tag}: config.env[{key!r}] looks like a real secret value; "
+                        "catalog entries must never ship secret values."
+                    )
+                if value != "" and not _PLACEHOLDER_RE.fullmatch(value):
+                    problems.append(
+                        f"{tag}: config.env[{key!r}] must be an empty string or a "
+                        "single <PLACEHOLDER> token -- the catalog declares which env "
+                        "vars a server needs, it never supplies their values."
+                    )
 
     return problems
 
@@ -2328,6 +2819,12 @@ def _validate_catalog_entry(idx: int, entry, seen_ids: set[str]) -> list[str]:
         tag = f"servers[{idx}] ({entry_id!r})"
         if not entry_id.isascii():
             problems.append(f"{tag}: 'id' must be ASCII (non-ASCII code points rejected).")
+        elif not _CATALOG_ID_RE.match(entry_id):
+            problems.append(
+                f"{tag}: 'id' must match ^[a-z0-9][a-z0-9._-]{{0,63}}$ "
+                "(it becomes an mcpServers JSON key and is interpolated into "
+                "Qt.AutoText widgets)."
+            )
         if entry_id in seen_ids:
             problems.append(f"{tag}: duplicate id.")
         seen_ids.add(entry_id)
@@ -2441,15 +2938,32 @@ def verify_catalog_signature(raw: bytes, sig: bytes, pubkeys: list[bytes]) -> bo
     return False
 
 
+def _verify_catalog_candidate(candidate: tuple[bytes, bytes] | None) -> tuple[dict | None, int]:
+    """Verify+load+validate one (raw, sig) candidate. Returns (None, -1) on any failure."""
+    if not candidate:
+        return None, -1
+    raw, sig = candidate
+    if not verify_catalog_signature(raw, sig, CATALOG_PUBKEYS):
+        return None, -1
+    try:
+        data = load_catalog(raw)
+    except (ValueError, TypeError):
+        return None, -1
+    if validate_catalog(data):
+        return None, -1
+    return data, catalog_version(data)
+
+
 def resolve_catalog(
     bundled: tuple[bytes, bytes] | None,
     cached: tuple[bytes, bytes] | None,
     remote: tuple[bytes, bytes] | None,
+    floor: int = 0,
 ) -> dict:
     """
-    Pick the highest-version catalog among bundled/cached/remote. Each
-    argument is either None (unavailable) or an (raw_bytes, signature_bytes)
-    pair.
+    Pick the highest-version catalog among bundled/cached/remote. Each of
+    bundled/cached/remote is either None (unavailable) or an (raw_bytes,
+    signature_bytes) pair.
 
     🔴 SECURITY: every candidate — including `bundled`, the copy frozen into
     this binary — is verified against CATALOG_PUBKEYS and re-validated from
@@ -2460,46 +2974,68 @@ def resolve_catalog(
     by virtue of being local. Signing (and checking the signature at
     runtime, every time) closes that.
 
-    Anti-rollback: a candidate's version is never accepted if it's lower
-    than the best verified candidate already found in this same resolution
-    pass — an attacker replaying an old, since-superseded signed catalog
-    can't downgrade you.
+    `floor` is a pure, caller-supplied lower bound (e.g. a persisted
+    "last accepted version" the GUI can load from disk and pass in) — this
+    function does no storage of its own.
 
-    Anti-freeze: a candidate whose version leaps more than
-    _CATALOG_MAX_VERSION_JUMP past the current best is also rejected. A
-    compromised/leaked signing key claiming an absurd future version would
-    otherwise permanently outrank every legitimate release from then on,
-    since the resolver always prefers the highest verified version — this
-    caps how far a single accepted jump can go.
+    Anti-rollback / anti-freeze, and WHY they apply to every candidate
+    including the first one evaluated: the previous version of this
+    function only ran these checks `if best_version >= 0`, i.e. once a
+    candidate had already been accepted in this pass. That let the FIRST
+    verified candidate through unconditionally — a signed catalog claiming
+    version=999999999 sailed straight past both guards if it happened to be
+    evaluated first, and rollback protection reset on every call anyway
+    (nothing persisted across restarts). Now both guards are anchored to
+    something that doesn't depend on iteration order:
+
+    - The anti-freeze cap is measured against the BUNDLED catalog's version
+      (verified independently, once), not against "whatever was accepted
+      so far in this loop." Bundled ships inside the binary, so it's the
+      one candidate that isn't attacker-supplied at resolve time — the
+      natural trust anchor. If bundled itself doesn't verify, `floor` is
+      the anchor instead.
+    - The anti-rollback floor is max(floor, bundled's version), so a
+      caller that persists `floor` across restarts gets real rollback
+      protection; a caller that doesn't still gets "never below bundled."
+
+    On a version TIE, the bundled candidate wins over cached/remote (it
+    previously lost ties to whichever candidate happened to be evaluated
+    last, silently preferring remote over bundled at equal version).
 
     Returns the winning catalog dict, or {} if nothing verified and
     validated.
     """
+    bundled_data, bundled_version = _verify_catalog_candidate(bundled)
+
+    anchor = bundled_version if bundled_version >= 0 else floor
+    min_accepted = max(floor, bundled_version if bundled_version >= 0 else 0)
+
+    candidates = (
+        ("bundled", bundled_data, bundled_version),
+        ("cached", *_verify_catalog_candidate(cached)),
+        ("remote", *_verify_catalog_candidate(remote)),
+    )
+
     best: dict = {}
     best_version = -1
+    best_is_bundled = False
 
-    for candidate in (bundled, cached, remote):
-        if not candidate:
-            continue
-        raw, sig = candidate
-        if not verify_catalog_signature(raw, sig, CATALOG_PUBKEYS):
-            continue
-        try:
-            data = load_catalog(raw)
-        except (ValueError, TypeError):
-            continue
-        if validate_catalog(data):
+    for source, data, version in candidates:
+        if data is None:
             continue
+        if version < min_accepted:
+            continue  # anti-rollback / below the persisted floor
+        if version > anchor + _CATALOG_MAX_VERSION_JUMP:
+            continue  # anti-freeze, capped against the bundled trust anchor
 
-        version = catalog_version(data)
-        if best_version >= 0:
-            if version < best_version:
-                continue  # anti-rollback
-            if version > best_version + _CATALOG_MAX_VERSION_JUMP:
-                continue  # anti-freeze
-
-        best = data
-        best_version = version
+        is_bundled = source == "bundled"
+        better = version > best_version or (
+            version == best_version and is_bundled and not best_is_bundled
+        )
+        if better:
+            best = data
+            best_version = version
+            best_is_bundled = is_bundled
 
     return best
 

catalog_review.py

@@ -0,0 +1,818 @@
+"""
+catalog_review.py -- pure, GUI-free review/diff/risk/crypto logic for the
+Catalog Console (issue #62).
+
+This module is deliberately Qt-free and network-free so every function in it
+is unit-testable offline, exactly like bcc_core.py. catalog_console.py (the
+PySide6 GUI) is a thin shell over these functions -- it owns Qt widgets,
+subprocess/git calls, and HTTP registry lookups; this module owns judgment.
+
+Nothing here is reimplemented from bcc_core: the command allowlist and the
+signature domain-separation prefix are imported, not retyped, so the two
+modules cannot silently drift apart (see bcc_core.validate_catalog /
+bcc_core.verify_catalog_signature and the project's "the check drifted on a
+new surface" recurring-bug lesson).
+"""
+
+from __future__ import annotations
+
+import os
+import re
+from collections.abc import Callable
+from dataclasses import dataclass, field
+from urllib.parse import urlsplit
+
+from bcc_core import _CATALOG_SIG_DOMAIN as CATALOG_SIG_DOMAIN
+from bcc_core import CATALOG_ALLOWED_COMMANDS
+from bcc_core import verify_catalog_signature as _verify_catalog_signature
+
+# --------------------------------------------------------------------------- #
+# Semantic diff
+# --------------------------------------------------------------------------- #
+
+# Top-level scalar/simple fields compared directly (not drilled into).
+_DIFF_FIELDS = (
+    "display",
+    "description",
+    "category",
+    "official",
+    "setup",
+    "homepage",
+    "docs_url",
+    "source",
+    "notes",
+    "stars",
+    "last_release",
+    "env_required",
+)
+
+
+@dataclass(frozen=True)
+class FieldChange:
+    """One field that differs between the old and new version of an entry."""
+
+    field: str
+    old: object
+    new: object
+
+
+@dataclass(frozen=True)
+class EntryChange:
+    """One catalog entry's change: added, removed, or changed.
+
+    `old`/`new` are the raw entry dicts (or None for added/removed) so risk
+    predicates and the GUI can inspect anything not captured by
+    `field_changes` (which only lists fields that actually differ).
+    """
+
+    entry_id: str
+    status: str  # "added" | "removed" | "changed"
+    old: dict | None
+    new: dict | None
+    field_changes: tuple[FieldChange, ...] = ()
+
+
+def _config_field_changes(old_cfg: dict | None, new_cfg: dict | None) -> list[FieldChange]:
+    old_cfg = old_cfg or {}
+    new_cfg = new_cfg or {}
+    changes: list[FieldChange] = []
+    for f in ("command", "args", "env"):
+        ov, nv = old_cfg.get(f), new_cfg.get(f)
+        if ov != nv:
+            changes.append(FieldChange(f"config.{f}", ov, nv))
+    return changes
+
+
+def _entry_field_changes(old_entry: dict, new_entry: dict) -> tuple[FieldChange, ...]:
+    changes: list[FieldChange] = []
+    for f in _DIFF_FIELDS:
+        ov, nv = old_entry.get(f), new_entry.get(f)
+        if ov != nv:
+            changes.append(FieldChange(f, ov, nv))
+    changes.extend(_config_field_changes(old_entry.get("config"), new_entry.get("config")))
+    return tuple(changes)
+
+
+def diff_catalogs(old: dict | None, new: dict | None) -> list[EntryChange]:
+    """Semantic (per-entry) diff between two parsed catalog dicts.
+
+    NOT a text diff: entries are matched by `id`, and each changed entry
+    reports exactly which fields differ (with before/after values), which is
+    what lets the Console render "command changed from X to Y" instead of a
+    JSON line diff a reviewer has to mentally reconstruct.
+
+    Entries missing/malformed `id` are ignored here -- that is a
+    validate_catalog() rejection, not a diffing concern, and diffing must not
+    silently invent a match for two differently-broken entries.
+    """
+    old_servers = {
+        e["id"]: e
+        for e in (old or {}).get("servers", []) or []
+        if isinstance(e, dict) and isinstance(e.get("id"), str) and e.get("id")
+    }
+    new_servers = {
+        e["id"]: e
+        for e in (new or {}).get("servers", []) or []
+        if isinstance(e, dict) and isinstance(e.get("id"), str) and e.get("id")
+    }
+
+    changes: list[EntryChange] = []
+    for entry_id in sorted(set(old_servers) | set(new_servers)):
+        old_e = old_servers.get(entry_id)
+        new_e = new_servers.get(entry_id)
+        if old_e is None:
+            changes.append(EntryChange(entry_id, "added", None, new_e, ()))
+        elif new_e is None:
+            changes.append(EntryChange(entry_id, "removed", old_e, None, ()))
+        elif old_e != new_e:
+            fc = _entry_field_changes(old_e, new_e)
+            if fc:
+                changes.append(EntryChange(entry_id, "changed", old_e, new_e, fc))
+    return changes
+
+
+# --------------------------------------------------------------------------- #
+# Risk annotations -- each predicate is pure and independently unit-tested.
+# --------------------------------------------------------------------------- #
+
+
+@dataclass(frozen=True)
+class RiskFinding:
+    severity: str  # "blocking" | "warning" | "info"
+    code: str
+    message: str
+
+
+def _escape_non_ascii(s: str) -> str:
+    """Render a string with any non-ASCII code point shown as an escape
+    sequence, so a homoglyph/RTL-override character can't visually pass as
+    the real thing in the review UI."""
+    return s.encode("unicode_escape").decode("ascii")
+
+
+def risk_env_required(change: EntryChange) -> list[RiskFinding]:
+    """A non-empty env_required value is blocking: catalog entries must ship
+    only the *names* of env vars the user fills in, never values."""
+    entry = change.new or {}
+    env_required = entry.get("env_required")
+    findings: list[RiskFinding] = []
+    if isinstance(env_required, dict):
+        for k, v in env_required.items():
+            if v not in (None, ""):
+                findings.append(
+                    RiskFinding(
+                        "blocking",
+                        "env_required_value",
+                        f"env_required[{k!r}] carries a non-empty value -- catalog "
+                        "entries must never ship secret values, only placeholder names.",
+                    )
+                )
+    return findings
+
+
+def risk_command_allowlist(change: EntryChange) -> list[RiskFinding]:
+    """A command outside bcc_core.CATALOG_ALLOWED_COMMANDS is blocking.
+    Imports the allowlist rather than redefining it."""
+    entry = change.new or {}
+    config = entry.get("config") or {}
+    command = config.get("command")
+    if isinstance(command, str) and command and command not in CATALOG_ALLOWED_COMMANDS:
+        return [
+            RiskFinding(
+                "blocking",
+                "command_not_allowed",
+                f"command {command!r} is not on the catalog allowlist "
+                f"({', '.join(sorted(CATALOG_ALLOWED_COMMANDS))}).",
+            )
+        ]
+    return []
+
+
+def risk_non_ascii(change: EntryChange) -> list[RiskFinding]:
+    """Non-ASCII code points in id/command/args are blocking -- homoglyph /
+    RTL-override typosquatting can make a malicious package name visually
+    identical to a legitimate one in a naive diff view."""
+    entry = change.new or {}
+    findings: list[RiskFinding] = []
+
+    entry_id = entry.get("id")
+    if isinstance(entry_id, str) and not entry_id.isascii():
+        findings.append(
+            RiskFinding(
+                "blocking",
+                "non_ascii_id",
+                f"id contains non-ASCII code points: {_escape_non_ascii(entry_id)!r}",
+            )
+        )
+
+    config = entry.get("config") or {}
+    command = config.get("command")
+    if isinstance(command, str) and not command.isascii():
+        findings.append(
+            RiskFinding(
+                "blocking",
+                "non_ascii_command",
+                f"command contains non-ASCII code points: {_escape_non_ascii(command)!r}",
+            )
+        )
+
+    for a in config.get("args") or []:
+        if isinstance(a, str) and not a.isascii():
+            findings.append(
+                RiskFinding(
+                    "blocking",
+                    "non_ascii_arg",
+                    f"arg contains non-ASCII code points: {_escape_non_ascii(a)!r}",
+                )
+            )
+
+    return findings
+
+
+def _npm_candidate_args(command: str | None, args: list[str]) -> list[str]:
+    if command != "npx":
+        return []
+    return [a for a in args if isinstance(a, str) and a and not a.startswith("-")]
+
+
+def split_npm_spec(spec: str) -> tuple[str, str | None]:
+    """Split an npm package spec into (name, version). version is None if
+    unpinned. Handles scoped (@scope/name@version) and unscoped
+    (name@version) specs."""
+    if spec.startswith("@"):
+        rest = spec[1:]
+        if "/" not in rest:
+            return spec, None  # malformed scope, can't tell -- treat unpinned
+        scope, _, remainder = rest.partition("/")
+        if "@" in remainder:
+            pkg_name, _, version = remainder.partition("@")
+            return f"@{scope}/{pkg_name}", (version or None)
+        return f"@{scope}/{remainder}", None
+    if "@" in spec:
+        name, _, version = spec.partition("@")
+        return name, (version or None)
+    return spec, None
+
+
+def is_pinned_npm_spec(spec: str) -> bool:
+    _name, version = split_npm_spec(spec)
+    return bool(version)
+
+
+_DOCKER_VALUE_FLAGS = {
+    "-e",
+    "--env",
+    "-v",
+    "--volume",
+    "-p",
+    "--publish",
+    "-w",
+    "--workdir",
+    "-u",
+    "--user",
+    "--name",
+    "--network",
+    "--entrypoint",
+}
+
+
+def docker_image_candidates(args: list[str]) -> list[str]:
+    """Best-effort extraction of the image reference from a `docker run
+    [OPTIONS] IMAGE [CMD...]` args list: the first positional token after
+    any leading `run` and flag(+value) pairs."""
+    candidates: list[str] = []
+    i = 0
+    while i < len(args):
+        a = args[i]
+        if a == "run":
+            i += 1
+            continue
+        if isinstance(a, str) and a.startswith("-"):
+            if "=" not in a and a in _DOCKER_VALUE_FLAGS:
+                i += 2
+                continue
+            i += 1
+            continue
+        if isinstance(a, str):
+            candidates.append(a)
+        break  # first positional token after `run` is the image ref
+    return candidates
+
+
+def is_pinned_docker_image(image: str) -> bool:
+    if "@sha256:" in image:
+        return True
+    tag_part = image.rsplit("/", 1)[-1]
+    if ":" not in tag_part:
+        return False  # no tag => implicit :latest
+    tag = tag_part.rsplit(":", 1)[-1]
+    return bool(tag) and tag != "latest"
+
+
+def risk_unpinned_package(change: EntryChange) -> list[RiskFinding]:
+    """Every entry must pin an exact version: an `@scope/pkg` npm arg with
+    no `@version`, or a docker image with no tag / `:latest`, is blocking.
+    A later-compromised package must not be able to auto-upgrade into every
+    user just because the catalog entry never pinned a version."""
+    entry = change.new or {}
+    config = entry.get("config") or {}
+    command = config.get("command")
+    args = config.get("args") or []
+    findings: list[RiskFinding] = []
+
+    if command == "npx":
+        for a in _npm_candidate_args(command, args):
+            if not is_pinned_npm_spec(a):
+                findings.append(
+                    RiskFinding(
+                        "blocking",
+                        "unpinned_npm_package",
+                        f"npm package arg {a!r} has no pinned @version.",
+                    )
+                )
+    elif command == "docker":
+        for img in docker_image_candidates(args):
+            if not is_pinned_docker_image(img):
+                findings.append(
+                    RiskFinding(
+                        "blocking",
+                        "unpinned_docker_image",
+                        f"docker image {img!r} is not pinned to an exact tag "
+                        "(uses :latest or no tag).",
+                    )
+                )
+
+    return findings
+
+
+_URL_FIELDS = ("homepage", "docs_url", "source")
+
+
+def _domain(url: str) -> str:
+    try:
+        return urlsplit(url).netloc.lower()
+    except ValueError:
+        return ""
+
+
+def risk_url_domain_change(change: EntryChange) -> list[RiskFinding]:
+    """Non-https URLs and, more importantly, a *domain change* on any URL
+    field are surfaced loudly with old-vs-new domains broken out -- the
+    lookalike-domain-swap defence."""
+    findings: list[RiskFinding] = []
+    old_entry = change.old or {}
+    new_entry = change.new or {}
+
+    for f in _URL_FIELDS:
+        new_url = new_entry.get(f)
+        if not isinstance(new_url, str) or not new_url:
+            continue
+        if not new_url.startswith("https://"):
+            findings.append(
+                RiskFinding("warning", "non_https_url", f"{f} is not https://: {new_url!r}")
+            )
+        old_url = old_entry.get(f)
+        if isinstance(old_url, str) and old_url:
+            old_domain, new_domain = _domain(old_url), _domain(new_url)
+            if old_domain and new_domain and old_domain != new_domain:
+                findings.append(
+                    RiskFinding(
+                        "warning",
+                        "domain_changed",
+                        f"{f} domain changed from {old_domain!r} to {new_domain!r} -- "
+                        "verify this isn't a lookalike-domain swap.",
+                    )
+                )
+    return findings
+
+
+def risk_new_entry(change: EntryChange) -> list[RiskFinding]:
+    """A brand-new entry is flagged for extra scrutiny -- not blocking on its
+    own, but it's the category of change the registry lookup exists for."""
+    if change.status == "added":
+        return [
+            RiskFinding(
+                "info",
+                "new_entry",
+                "Brand-new catalog entry -- extra scrutiny: check publisher identity "
+                "via the registry lookup before signing.",
+            )
+        ]
+    return []
+
+
+_RISK_PREDICATES: tuple[Callable[[EntryChange], list[RiskFinding]], ...] = (
+    risk_env_required,
+    risk_command_allowlist,
+    risk_non_ascii,
+    risk_unpinned_package,
+    risk_url_domain_change,
+    risk_new_entry,
+)
+
+
+def entry_risk_findings(change: EntryChange) -> list[RiskFinding]:
+    """Run every risk predicate against one entry change and return the
+    combined findings (order matches _RISK_PREDICATES)."""
+    findings: list[RiskFinding] = []
+    for predicate in _RISK_PREDICATES:
+        findings.extend(predicate(change))
+    return findings
+
+
+def has_blocking_risk(change: EntryChange) -> bool:
+    return any(f.severity == "blocking" for f in entry_risk_findings(change))
+
+
+# --------------------------------------------------------------------------- #
+# Review session: acknowledge-gating + TOCTOU blob-SHA pinning
+# --------------------------------------------------------------------------- #
+
+
+@dataclass
+class ReviewSession:
+    """State for one review pass. `pinned_blob_sha` is the git blob SHA of
+    data/catalog.json as it existed the moment review began -- see
+    can_sign()/sign_precondition().
+
+    `loaded_ref` is the exact ref this review was loaded from ("main", or a
+    PR's `refs/pull/<n>/head`) -- see issue #68 finding 1. It exists so the
+    Sign path can re-resolve the TOCTOU blob SHA from *the ref that was
+    actually reviewed*, instead of a hardcoded "main" that silently diverges
+    from the reviewed ref on every PR review (the bug that made the PR path
+    unable to sign at all, and forced everyone onto the vacuous
+    main-vs-itself path instead).
+    """
+
+    pinned_blob_sha: str
+    old_catalog: dict
+    new_catalog: dict
+    loaded_ref: str = "main"
+    changes: list[EntryChange] = field(default_factory=list)
+    acknowledged: set[str] = field(default_factory=set)
+
+    def __post_init__(self) -> None:
+        if not self.changes:
+            self.changes = diff_catalogs(self.old_catalog, self.new_catalog)
+
+
+def start_review(
+    pinned_blob_sha: str,
+    old_catalog: dict,
+    new_catalog: dict,
+    loaded_ref: str = "main",
+) -> ReviewSession:
+    return ReviewSession(
+        pinned_blob_sha=pinned_blob_sha,
+        old_catalog=old_catalog,
+        new_catalog=new_catalog,
+        loaded_ref=loaded_ref,
+    )
+
+
+def find_last_signed_catalog_raw(
+    candidates: list[bytes], sig: bytes, pubkeys: list[bytes]
+) -> bytes | None:
+    """Given `candidates` (candidate raw catalog.json byte-strings -- e.g.
+    successive historical versions from git log, most-recent-first),
+    return the first one whose signature verifies against `sig`/`pubkeys`,
+    or None if none do.
+
+    This is how source="main" review diffs against "the last catalog a
+    maintainer actually signed" instead of against itself (issue #68
+    finding 1): `catalog_console.last_signed_catalog_raw` walks
+    data/catalog.json's git history on main and hands the candidates here.
+    """
+    for raw in candidates:
+        if _verify_catalog_signature(raw, sig, pubkeys):
+            return raw
+    return None
+
+
+def acknowledge_entry(session: ReviewSession, entry_id: str) -> None:
+    ids = {c.entry_id for c in session.changes}
+    if entry_id not in ids:
+        raise ValueError(f"{entry_id!r} is not part of this review session's diff.")
+    session.acknowledged.add(entry_id)
+
+
+def unacknowledge_entry(session: ReviewSession, entry_id: str) -> None:
+    session.acknowledged.discard(entry_id)
+
+
+def all_entries_acknowledged(session: ReviewSession) -> bool:
+    return {c.entry_id for c in session.changes} <= session.acknowledged
+
+
+# NOTE for future editors: do NOT add an "acknowledge all" shortcut here, now
+# or ever. The friction of individually acknowledging every changed entry is
+# the entire point of this tool (issue #62) -- a shortcut would let a tired
+# reviewer rubber-stamp a diff exactly like the "merge PR, run script, push"
+# reflex this Console exists to replace. If this comment is the only thing
+# stopping you, that is the point: it is stopping you on purpose.
+
+
+@dataclass(frozen=True)
+class SignDecision:
+    ok: bool
+    reason: str | None = None
+
+
+def can_sign(session: ReviewSession, current_blob_sha: str) -> SignDecision:
+    """Whether the Sign button may fire right now.
+
+    Four independent gates, all required, checked in this order:
+
+    0. The diff must be non-empty. An empty diff historically meant "Sign
+       unlocks instantly" (`set() <= set()` is vacuously True), which is
+       exactly backwards: a vacuously-satisfied gate is worse than no gate
+       at all, because it *manufactures confidence* -- the signature looks
+       identical to one produced by a real review. "Nothing changed" must
+       mean "nothing to sign", never "sign unlocked". (Issue #68 finding 1;
+       this is what let commit b08cf21 sign all 19 entries with zero of them
+       ever reviewed.)
+    1. TOCTOU: `current_blob_sha` (fetched fresh, immediately before signing,
+       from the ref that was actually reviewed -- see sign_precondition())
+       must match the blob SHA pinned when review began. If the bytes on the
+       remote changed since -- a new commit pushed to the same PR, a
+       force-push, another PR merged in between -- signing is refused and a
+       re-review is forced. This is what makes "signing is the approval act"
+       true rather than aspirational: the signature is bound to the exact
+       reviewed bytes, not to "whatever the file happens to be now".
+    2. No blocking risk finding may be outstanding on ANY changed entry, full
+       stop -- checked here, not just in the GUI. The GUI additionally
+       disables the acknowledge checkbox for a blocking entry, but that is a
+       UI nicety, not the enforcement point: if this pure gate didn't also
+       check it, a blocking risk would only be stopped by the GUI happening
+       to have wired the checkbox correctly, and nothing would catch a
+       regression in that wiring. The GUI must not be the only thing
+       standing between a blocking risk and a signature.
+    3. Every changed entry in the diff must be individually acknowledged.
+    """
+    if not session.changes:
+        return SignDecision(
+            False,
+            "Nothing to sign: this review's diff is empty. If you expected "
+            "changes here, you may be diffing the wrong source/ref.",
+        )
+    if current_blob_sha != session.pinned_blob_sha:
+        return SignDecision(
+            False,
+            "The reviewed bytes changed since this review began (blob SHA "
+            "mismatch) -- re-review required before signing.",
+        )
+    blocking_ids = sorted({c.entry_id for c in session.changes if has_blocking_risk(c)})
+    if blocking_ids:
+        return SignDecision(
+            False,
+            "Blocking risk finding(s) outstanding on: "
+            f"{', '.join(blocking_ids)} -- fix the underlying change, do not sign around it.",
+        )
+    if not all_entries_acknowledged(session):
+        pending = sorted({c.entry_id for c in session.changes} - session.acknowledged)
+        return SignDecision(
+            False, f"Not every changed entry has been acknowledged yet: {', '.join(pending)}"
+        )
+    return SignDecision(True, None)
+
+
+def sign_precondition(
+    session: ReviewSession, resolve_blob_sha: Callable[[str], str]
+) -> SignDecision:
+    """The real Sign-button gate: resolves the current TOCTOU blob SHA from
+    *the ref this session was actually loaded from* (`session.loaded_ref`),
+    never a hardcoded "main", then delegates to can_sign().
+
+    `resolve_blob_sha` is injected so this stays testable without git/Qt --
+    catalog_console.ReviewWindow._on_sign passes a real resolver
+    (fetch_ref + blob_sha_at against self.repo_dir); tests pass a fake
+    dict-backed lookup. This is the fix for issue #68 finding 1's first bug:
+    `_on_sign` used to hardcode `fetch_ref(self.repo_dir, "main")` as the
+    comparison ref, so for any PR review (where `loaded_ref` is the PR's
+    head, not main) the SHAs differed by definition and Sign could never
+    fire -- and the retry path re-called the same hardcoded resolver, so it
+    re-pinned the same wrong value and looped forever instead of forcing a
+    genuine re-review.
+    """
+    current_blob_sha = resolve_blob_sha(session.loaded_ref)
+    return can_sign(session, current_blob_sha)
+
+
+def catalog_signing_message(raw_bytes: bytes) -> bytes:
+    """The exact bytes that get signed: bcc_core's domain-separation prefix
+    (imported, never retyped) + the raw catalog bytes. Using this function
+    guarantees the Console's signature and bcc_core.verify_catalog_signature
+    can never drift apart on the prefix."""
+    return CATALOG_SIG_DOMAIN + raw_bytes
+
+
+# --------------------------------------------------------------------------- #
+# Key management: passphrase-encrypted-at-rest Ed25519 seed
+#
+# The private key is NEVER stored plaintext, never an env var, never
+# committed. encrypt_private_key/decrypt_private_key are pure and offline
+# (scrypt KDF + AES-256-GCM via `cryptography`, already a project
+# dependency); catalog_console.py decides WHERE the resulting blob lives
+# (OS keychain if available, else a file outside the repo).
+# --------------------------------------------------------------------------- #
+
+_KDF_SALT_LEN = 16
+_KDF_N = 2**15  # scrypt cost parameter, tuned for a one-off interactive unlock
+_KDF_R = 8
+_KDF_P = 1
+_NONCE_LEN = 12
+_AAD = b"bcc-catalog-console-key-v1"
+
+
+def _derive_key(passphrase: str, salt: bytes) -> bytes:
+    from cryptography.hazmat.primitives.kdf.scrypt import Scrypt
+
+    kdf = Scrypt(salt=salt, length=32, n=_KDF_N, r=_KDF_R, p=_KDF_P)
+    return kdf.derive(passphrase.encode("utf-8"))
+
+
+def generate_keypair() -> tuple[bytes, bytes]:
+    """Generate a new Ed25519 keypair. Returns (seed_32_bytes, pubkey_32_bytes)."""
+    from cryptography.hazmat.primitives.asymmetric.ed25519 import Ed25519PrivateKey
+    from cryptography.hazmat.primitives.serialization import Encoding, PublicFormat
+
+    private_key = Ed25519PrivateKey.generate()
+    seed = private_key.private_bytes_raw()
+    pubkey = private_key.public_key().public_bytes(Encoding.Raw, PublicFormat.Raw)
+    return seed, pubkey
+
+
+def encrypt_private_key(seed: bytes, passphrase: str) -> bytes:
+    """Encrypt a 32-byte Ed25519 seed at rest with a passphrase. Returns a
+    self-contained blob: salt || nonce || ciphertext+tag."""
+    from cryptography.hazmat.primitives.ciphers.aead import AESGCM
+
+    if len(seed) != 32:
+        raise ValueError(f"expected a 32-byte raw Ed25519 seed, got {len(seed)} bytes")
+    if not passphrase:
+        raise ValueError("a non-empty passphrase is required")
+    salt = os.urandom(_KDF_SALT_LEN)
+    key = _derive_key(passphrase, salt)
+    nonce = os.urandom(_NONCE_LEN)
+    ciphertext = AESGCM(key).encrypt(nonce, seed, _AAD)
+    return salt + nonce + ciphertext
+
+
+def decrypt_private_key(blob: bytes, passphrase: str) -> bytes:
+    """Decrypt a blob produced by encrypt_private_key. Raises ValueError on a
+    wrong passphrase or corrupt blob -- never silently returns garbage."""
+    from cryptography.exceptions import InvalidTag
+    from cryptography.hazmat.primitives.ciphers.aead import AESGCM
+
+    if len(blob) < _KDF_SALT_LEN + _NONCE_LEN:
+        raise ValueError("key blob is too short to be valid")
+    salt = blob[:_KDF_SALT_LEN]
+    nonce = blob[_KDF_SALT_LEN : _KDF_SALT_LEN + _NONCE_LEN]
+    ciphertext = blob[_KDF_SALT_LEN + _NONCE_LEN :]
+    key = _derive_key(passphrase, salt)
+    try:
+        return AESGCM(key).decrypt(nonce, ciphertext, _AAD)
+    except InvalidTag as e:
+        raise ValueError("wrong passphrase or corrupted key file") from e
+
+
+def sign_catalog_bytes(raw: bytes, seed: bytes) -> bytes:
+    """Sign `raw` catalog bytes with a 32-byte Ed25519 seed, using the exact
+    domain-separated message bcc_core.verify_catalog_signature expects."""
+    from cryptography.hazmat.primitives.asymmetric.ed25519 import Ed25519PrivateKey
+
+    if len(seed) != 32:
+        raise ValueError(f"expected a 32-byte raw Ed25519 seed, got {len(seed)} bytes")
+    private_key = Ed25519PrivateKey.from_private_bytes(seed)
+    return private_key.sign(catalog_signing_message(raw))
+
+
+# --------------------------------------------------------------------------- #
+# Registry lookup -- the check a human genuinely can't do.
+#
+# The network call itself is injected (a `Fetcher` callable) so this stays
+# testable offline; catalog_console.py supplies the real npm/PyPI HTTP
+# fetcher. Fails soft everywhere: a fetcher returning None/raising just
+# yields RegistryInfo(available=False), never an exception into the caller
+# and never a block on review.
+# --------------------------------------------------------------------------- #
+
+
+@dataclass(frozen=True)
+class PackageRef:
+    entry_id: str
+    ecosystem: str  # "npm" | "pypi"
+    name: str
+    version: str | None
+
+
+def split_pypi_spec(spec: str) -> tuple[str, str | None]:
+    for sep in ("==", "@"):
+        if sep in spec:
+            name, _, version = spec.partition(sep)
+            return name, (version or None)
+    return spec, None
+
+
+def extract_package_refs(entry: dict) -> list[PackageRef]:
+    """Pull out the package(s) a basic-tier entry's args reference, for the
+    registry lookup. Returns [] for link-only entries or entries whose
+    command isn't npx/uvx (docker images aren't registry-lookup candidates
+    in the npm/PyPI sense used here)."""
+    config = entry.get("config") or {}
+    command = config.get("command")
+    args = config.get("args") or []
+    entry_id = entry.get("id", "") if isinstance(entry.get("id"), str) else ""
+    refs: list[PackageRef] = []
+
+    if command == "npx":
+        for a in _npm_candidate_args(command, args):
+            name, version = split_npm_spec(a)
+            if name:
+                refs.append(PackageRef(entry_id, "npm", name, version))
+    elif command == "uvx":
+        for a in args:
+            if isinstance(a, str) and a and not a.startswith("-"):
+                name, version = split_pypi_spec(a)
+                if name:
+                    refs.append(PackageRef(entry_id, "pypi", name, version))
+                break  # `uvx <pkg>` -- first positional token is the package
+
+    return refs
+
+
+@dataclass(frozen=True)
+class RegistryInfo:
+    ref: PackageRef
+    available: bool
+    publisher: str | None = None
+    age_days: int | None = None
+    last_release: str | None = None
+    downloads: int | None = None
+    near_neighbor_ids: tuple[str, ...] = ()
+
+
+Fetcher = Callable[[PackageRef], dict | None]
+
+
+def edit_distance(a: str, b: str) -> int:
+    """Levenshtein distance, iterative DP (no recursion depth concerns)."""
+    if a == b:
+        return 0
+    la, lb = len(a), len(b)
+    if la == 0:
+        return lb
+    if lb == 0:
+        return la
+    prev = list(range(lb + 1))
+    for i, ca in enumerate(a, 1):
+        cur = [i] + [0] * lb
+        for j, cb in enumerate(b, 1):
+            cost = 0 if ca == cb else 1
+            cur[j] = min(prev[j] + 1, cur[j - 1] + 1, prev[j - 1] + cost)
+        prev = cur
+    return prev[lb]
+
+
+def near_neighbor_ids(name: str, other_ids: list[str], max_distance: int = 2) -> list[str]:
+    """Catalog ids within `max_distance` edits of `name` (case-insensitive),
+    excluding an exact match -- the dependency-confusion / typosquat
+    near-neighbour warning."""
+    lname = name.lower()
+    return [
+        oid
+        for oid in other_ids
+        if oid != name and edit_distance(lname, oid.lower()) <= max_distance
+    ]
+
+
+def lookup_registry_info(
+    ref: PackageRef, fetcher: Fetcher, all_entry_ids: list[str]
+) -> RegistryInfo:
+    """Resolve one package against the live registry via the injected
+    fetcher. Never raises: any fetcher exception or falsy return means
+    `available=False` ("unavailable"), which the GUI renders plainly rather
+    than blocking or erroring the review."""
+    neighbors = tuple(near_neighbor_ids(ref.name, all_entry_ids))
+    try:
+        raw = fetcher(ref)
+    except Exception:
+        raw = None
+    if not raw:
+        return RegistryInfo(ref=ref, available=False, near_neighbor_ids=neighbors)
+    return RegistryInfo(
+        ref=ref,
+        available=True,
+        publisher=raw.get("publisher"),
+        age_days=raw.get("age_days"),
+        last_release=raw.get("last_release"),
+        downloads=raw.get("downloads"),
+        near_neighbor_ids=neighbors,
+    )
+
+
+_NON_ASCII_RE = re.compile(r"[^\x00-\x7f]")
+
+
+def contains_non_ascii(s: str) -> bool:
+    return bool(_NON_ASCII_RE.search(s))

data/catalog.json.sig

@@ -0,0 +1,2 @@
+ы<8¶ђt2ішл„»‰/НЕ0Тjcw&`
+тrH«MўК›єrBL,0AS€!Х2–иже.SТ°ч–'Agm

scripts/sign_checksums.py

@@ -42,6 +42,25 @@ from pathlib import Path
 # message signed by the same key.
 DOMAIN_PREFIX = b"bcc-release-v1|"
 
+# Public half of the RELEASE signing key(s) -- a SEPARATE keypair from
+# bcc_core.CATALOG_PUBKEYS (issue #68 finding 5). The catalog key is the
+# offline, Console-only root of trust for what BCC executes; this key is
+# CI-resident and signs ONLY the release SHA256SUMS manifest, never the
+# catalog. Keeping them apart means a CI/repo-secret compromise burns the
+# release key -- annoying, but it never lets an attacker sign a catalog a
+# user's binary would trust. A LIST (not a single key), mirroring
+# CATALOG_PUBKEYS, so the release key can be rotated without invalidating
+# the signature on every past release: verification accepts a match against
+# ANY key here.
+#
+# Empty until the maintainer generates the release keypair (separately from
+# the catalog keypair) and pastes the public half in:
+#     python catalog_console.py keygen --release
+# This is intentionally NOT pre-populated with a placeholder that looks
+# like a real key -- release.yml's signing-smoke-test fails closed (loudly)
+# on an empty list rather than silently verifying against nothing.
+RELEASE_PUBKEYS: list[bytes] = []
+
 CHUNK_SIZE = 1024 * 1024
 
 
@@ -135,6 +154,19 @@ def public_key_b64_from_seed(seed_b64: str) -> str:
     return base64.b64encode(raw).decode("ascii")
 
 
+def verify_checksums_against_any(pubkeys: list[bytes], sums_text: str, signature: bytes) -> bool:
+    """Verify `signature` against ANY key in `pubkeys` (each a raw 32-byte
+    Ed25519 public key). Mirrors bcc_core.verify_catalog_signature's
+    rotation-friendly "any currently-trusted key" semantics, applied to
+    RELEASE_PUBKEYS instead of the catalog's key list. Returns False (never
+    raises) for an empty `pubkeys` list -- fails closed rather than
+    vacuously verifying against nothing."""
+    return any(
+        verify_checksums(base64.b64encode(pk).decode("ascii"), sums_text, signature)
+        for pk in pubkeys
+    )
+
+
 # --------------------------------------------------------------------------- #
 # CLI
 # --------------------------------------------------------------------------- #

tests/test_catalog_console_packaging.py

@@ -0,0 +1,66 @@
+"""Asserts the maintainer-only Catalog Console (catalog_console.py,
+catalog_review.py) is never bundled into the release binary.
+
+A signing/review tool shipping to end users would be an own-goal (issue
+#62): it has no reason to run on a user's machine, and its presence would
+be a confusing artefact of a build that's supposed to be a thin GUI over
+mcpServers config editing."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+REPO_ROOT = Path(__file__).resolve().parent.parent
+SPEC_PATH = REPO_ROOT / "bcc.spec"
+
+_EXCLUDED_FILES = ("catalog_console.py", "catalog_review.py")
+
+
+def test_spec_file_exists():
+    assert SPEC_PATH.exists()
+
+
+def test_console_files_not_named_in_spec():
+    """The spec text must never reference either maintainer-only module --
+    not as the Analysis entry point, not in datas, not anywhere."""
+    spec_text = SPEC_PATH.read_text(encoding="utf-8")
+    for filename in _EXCLUDED_FILES:
+        assert filename not in spec_text, (
+            f"{filename} must never be referenced by bcc.spec -- it is a "
+            "maintainer-only tool and must not ship to users."
+        )
+
+
+def test_spec_analysis_entry_point_is_bcc_py_only():
+    """PyInstaller's Analysis(...) call determines the dependency-scanned
+    entry point(s); it must be bcc.py alone."""
+    spec_text = SPEC_PATH.read_text(encoding="utf-8")
+    assert 'Analysis(\n    ["bcc.py"],' in spec_text or 'Analysis(["bcc.py"]' in spec_text, (
+        "bcc.spec's Analysis(...) entry point changed shape -- re-verify by hand "
+        "that catalog_console.py / catalog_review.py are still excluded."
+    )
+
+
+def test_console_modules_exist_but_are_standalone_top_level_files():
+    """Sanity check the files this test is guarding actually exist as
+    top-level modules (not, say, silently moved into a package PyInstaller's
+    Analysis would still pick up as an implicit import of bcc.py)."""
+    for filename in _EXCLUDED_FILES:
+        assert (REPO_ROOT / filename).exists()
+        # bcc.py must not import them.
+        bcc_text = (REPO_ROOT / "bcc.py").read_text(encoding="utf-8")
+        module_name = filename.removesuffix(".py")
+        assert f"import {module_name}" not in bcc_text
+        assert f"from {module_name}" not in bcc_text
+
+
+def test_requirements_files_do_not_reference_console_only_needs():
+    """catalog_console.py's only import beyond the shipped stack is the
+    optional `keyring` package, which is intentionally NOT added as a hard
+    dependency anywhere a user install would pick it up."""
+    for req_file in ("requirements.txt", "requirements-dev.txt"):
+        path = REPO_ROOT / req_file
+        if not path.exists():
+            continue
+        text = path.read_text(encoding="utf-8").lower()
+        assert "keyring" not in text

tests/test_catalog_review.py

@@ -0,0 +1,695 @@
+"""Tests for catalog_review.py -- semantic diff, risk predicates, review
+session (acknowledge-gating + TOCTOU blob pinning), key encryption, and
+registry-lookup logic for the Catalog Console (issue #62)."""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+
+import bcc_core as c
+import catalog_review as r
+
+
+def _entry(**overrides):
+    base = {
+        "id": "filesystem",
+        "display": "Filesystem",
+        "description": "desc",
+        "category": "files",
+        "homepage": "https://github.com/modelcontextprotocol/servers",
+        "stars": 100,
+        "official": True,
+        "setup": "basic",
+        "config": {
+            "command": "npx",
+            "args": ["-y", "@modelcontextprotocol/server-filesystem@1.0.0"],
+        },
+        "env_required": {},
+        "docs_url": "https://github.com/modelcontextprotocol/servers",
+        "notes": "",
+        "last_release": "2026-01-01",
+    }
+    base.update(overrides)
+    return base
+
+
+def _catalog(*entries):
+    return {"schema": 1, "version": 1, "servers": list(entries)}
+
+
+# --------------------------------------------------------------------------- #
+# diff_catalogs
+# --------------------------------------------------------------------------- #
+def test_diff_detects_added_entry():
+    old = _catalog()
+    new = _catalog(_entry())
+    changes = r.diff_catalogs(old, new)
+    assert len(changes) == 1
+    assert changes[0].status == "added"
+    assert changes[0].entry_id == "filesystem"
+    assert changes[0].old is None
+
+
+def test_diff_detects_removed_entry():
+    old = _catalog(_entry())
+    new = _catalog()
+    changes = r.diff_catalogs(old, new)
+    assert len(changes) == 1
+    assert changes[0].status == "removed"
+    assert changes[0].new is None
+
+
+def test_diff_detects_no_change():
+    e = _entry()
+    old = _catalog(e)
+    new = _catalog(dict(e))
+    assert r.diff_catalogs(old, new) == []
+
+
+def test_diff_detects_changed_command_and_args():
+    old = _catalog(_entry())
+    new = _catalog(_entry(config={"command": "uvx", "args": ["other-pkg@2.0.0"]}))
+    changes = r.diff_catalogs(old, new)
+    assert len(changes) == 1
+    ch = changes[0]
+    assert ch.status == "changed"
+    fields = {fc.field for fc in ch.field_changes}
+    assert "config.command" in fields
+    assert "config.args" in fields
+
+
+def test_diff_detects_description_change():
+    old = _catalog(_entry())
+    new = _catalog(_entry(description="new description"))
+    changes = r.diff_catalogs(old, new)
+    assert changes[0].field_changes == (r.FieldChange("description", "desc", "new description"),)
+
+
+def test_diff_ignores_entries_without_id():
+    old = _catalog()
+    new = _catalog({"display": "no id"})
+    assert r.diff_catalogs(old, new) == []
+
+
+def test_diff_multiple_entries_sorted_by_id():
+    old = _catalog(_entry(id="zeta"), _entry(id="alpha"))
+    new = _catalog(
+        _entry(id="zeta", description="changed"), _entry(id="alpha", description="changed")
+    )
+    changes = r.diff_catalogs(old, new)
+    assert [c_.entry_id for c_ in changes] == ["alpha", "zeta"]
+
+
+# --------------------------------------------------------------------------- #
+# risk_env_required
+# --------------------------------------------------------------------------- #
+def test_risk_env_required_blocking_on_nonempty_value():
+    change = r.EntryChange("x", "changed", None, _entry(env_required={"API_KEY": "sk-real-value"}))
+    findings = r.risk_env_required(change)
+    assert len(findings) == 1
+    assert findings[0].severity == "blocking"
+    assert findings[0].code == "env_required_value"
+
+
+def test_risk_env_required_clean_on_empty_value():
+    change = r.EntryChange("x", "changed", None, _entry(env_required={"API_KEY": ""}))
+    assert r.risk_env_required(change) == []
+
+
+# --------------------------------------------------------------------------- #
+# risk_command_allowlist
+# --------------------------------------------------------------------------- #
+def test_risk_command_allowlist_blocks_disallowed_command():
+    change = r.EntryChange(
+        "x", "changed", None, _entry(config={"command": "bash", "args": ["-c", "evil"]})
+    )
+    findings = r.risk_command_allowlist(change)
+    assert len(findings) == 1
+    assert findings[0].severity == "blocking"
+
+
+def test_risk_command_allowlist_allows_listed_command():
+    for cmd in sorted(c.CATALOG_ALLOWED_COMMANDS):
+        change = r.EntryChange("x", "changed", None, _entry(config={"command": cmd, "args": []}))
+        assert r.risk_command_allowlist(change) == []
+
+
+# --------------------------------------------------------------------------- #
+# risk_non_ascii
+# --------------------------------------------------------------------------- #
+def test_risk_non_ascii_flags_homoglyph_id():
+    # Cyrillic 'а' (U+0430) instead of Latin 'a' -- classic homoglyph swap.
+    evil_id = "filаsystem"
+    change = r.EntryChange(evil_id, "changed", None, _entry(id=evil_id))
+    findings = r.risk_non_ascii(change)
+    assert any(f.code == "non_ascii_id" for f in findings)
+    assert findings[0].severity == "blocking"
+    # the offending string must be rendered with escapes, not raw
+    assert "\\u0430" in findings[0].message
+
+
+def test_risk_non_ascii_flags_arg():
+    change = r.EntryChange(
+        "x", "changed", None, _entry(config={"command": "npx", "args": ["pаckage@1.0.0"]})
+    )
+    findings = r.risk_non_ascii(change)
+    assert any(f.code == "non_ascii_arg" for f in findings)
+
+
+def test_risk_non_ascii_clean_for_ascii_entry():
+    change = r.EntryChange("x", "changed", None, _entry())
+    assert r.risk_non_ascii(change) == []
+
+
+# --------------------------------------------------------------------------- #
+# risk_unpinned_package
+# --------------------------------------------------------------------------- #
+def test_risk_unpinned_npm_package_no_version():
+    change = r.EntryChange(
+        "x",
+        "changed",
+        None,
+        _entry(config={"command": "npx", "args": ["-y", "@scope/pkg"]}),
+    )
+    findings = r.risk_unpinned_package(change)
+    assert len(findings) == 1
+    assert findings[0].code == "unpinned_npm_package"
+    assert findings[0].severity == "blocking"
+
+
+def test_risk_pinned_npm_package_is_clean():
+    change = r.EntryChange(
+        "x",
+        "changed",
+        None,
+        _entry(config={"command": "npx", "args": ["-y", "@scope/pkg@1.2.3"]}),
+    )
+    assert r.risk_unpinned_package(change) == []
+
+
+def test_risk_unpinned_unscoped_npm_package():
+    change = r.EntryChange(
+        "x", "changed", None, _entry(config={"command": "npx", "args": ["-y", "somepkg"]})
+    )
+    findings = r.risk_unpinned_package(change)
+    assert len(findings) == 1
+
+
+def test_risk_unpinned_docker_latest_tag():
+    change = r.EntryChange(
+        "x",
+        "changed",
+        None,
+        _entry(config={"command": "docker", "args": ["run", "-i", "--rm", "myimage:latest"]}),
+    )
+    findings = r.risk_unpinned_package(change)
+    assert len(findings) == 1
+    assert findings[0].code == "unpinned_docker_image"
+
+
+def test_risk_unpinned_docker_no_tag():
+    change = r.EntryChange(
+        "x", "changed", None, _entry(config={"command": "docker", "args": ["run", "myimage"]})
+    )
+    findings = r.risk_unpinned_package(change)
+    assert len(findings) == 1
+
+
+def test_risk_pinned_docker_image_is_clean():
+    change = r.EntryChange(
+        "x",
+        "changed",
+        None,
+        _entry(config={"command": "docker", "args": ["run", "-i", "--rm", "myimage:1.2.3"]}),
+    )
+    assert r.risk_unpinned_package(change) == []
+
+
+def test_risk_docker_digest_pin_is_clean():
+    change = r.EntryChange(
+        "x",
+        "changed",
+        None,
+        _entry(
+            config={
+                "command": "docker",
+                "args": ["run", "myimage@sha256:" + "a" * 64],
+            }
+        ),
+    )
+    assert r.risk_unpinned_package(change) == []
+
+
+# --------------------------------------------------------------------------- #
+# risk_url_domain_change
+# --------------------------------------------------------------------------- #
+def test_risk_url_domain_change_warns_on_lookalike_swap():
+    old_entry = _entry(homepage="https://github.com/foo/bar")
+    new_entry = _entry(homepage="https://githu6.com/foo/bar")
+    change = r.EntryChange("x", "changed", old_entry, new_entry)
+    findings = r.risk_url_domain_change(change)
+    assert any(f.code == "domain_changed" for f in findings)
+    domain_finding = next(f for f in findings if f.code == "domain_changed")
+    assert "github.com" in domain_finding.message
+    assert "githu6.com" in domain_finding.message
+    assert domain_finding.severity == "warning"
+
+
+def test_risk_url_domain_change_clean_when_domain_unchanged():
+    old_entry = _entry(homepage="https://github.com/foo/bar")
+    new_entry = _entry(homepage="https://github.com/foo/bar-renamed")
+    change = r.EntryChange("x", "changed", old_entry, new_entry)
+    assert r.risk_url_domain_change(change) == []
+
+
+def test_risk_url_non_https_warns():
+    new_entry = _entry(homepage="http://example.com")
+    change = r.EntryChange("x", "changed", None, new_entry)
+    findings = r.risk_url_domain_change(change)
+    assert any(f.code == "non_https_url" for f in findings)
+
+
+# --------------------------------------------------------------------------- #
+# risk_new_entry / entry_risk_findings / has_blocking_risk
+# --------------------------------------------------------------------------- #
+def test_risk_new_entry_flags_added():
+    change = r.EntryChange("x", "added", None, _entry())
+    findings = r.risk_new_entry(change)
+    assert len(findings) == 1
+    assert findings[0].severity == "info"
+
+
+def test_risk_new_entry_silent_for_changed():
+    change = r.EntryChange("x", "changed", _entry(), _entry(description="x"))
+    assert r.risk_new_entry(change) == []
+
+
+def test_has_blocking_risk_true_for_disallowed_command():
+    change = r.EntryChange("x", "changed", None, _entry(config={"command": "bash", "args": []}))
+    assert r.has_blocking_risk(change) is True
+
+
+def test_has_blocking_risk_false_for_clean_entry():
+    change = r.EntryChange("x", "changed", _entry(), _entry(description="new"))
+    assert r.has_blocking_risk(change) is False
+
+
+# --------------------------------------------------------------------------- #
+# ReviewSession: acknowledge gating
+# --------------------------------------------------------------------------- #
+def test_start_review_computes_diff():
+    old = _catalog()
+    new = _catalog(_entry())
+    session = r.start_review("sha1", old, new)
+    assert len(session.changes) == 1
+
+
+def test_all_entries_acknowledged_false_initially():
+    session = r.start_review("sha1", _catalog(), _catalog(_entry()))
+    assert r.all_entries_acknowledged(session) is False
+
+
+def test_acknowledge_entry_marks_acknowledged():
+    session = r.start_review("sha1", _catalog(), _catalog(_entry()))
+    r.acknowledge_entry(session, "filesystem")
+    assert r.all_entries_acknowledged(session) is True
+
+
+def test_acknowledge_unknown_entry_raises():
+    session = r.start_review("sha1", _catalog(), _catalog(_entry()))
+    with pytest.raises(ValueError):
+        r.acknowledge_entry(session, "not-in-diff")
+
+
+def test_acknowledge_gating_requires_every_entry():
+    session = r.start_review("sha1", _catalog(), _catalog(_entry(id="a"), _entry(id="b")))
+    r.acknowledge_entry(session, "a")
+    assert r.all_entries_acknowledged(session) is False
+    r.acknowledge_entry(session, "b")
+    assert r.all_entries_acknowledged(session) is True
+
+
+def test_no_acknowledge_all_shortcut_and_gate_is_real():
+    """Two things, both load-bearing (issue #68: the original version of
+    this test asserted ONLY the first half, and passed the entire time the
+    gate below it was vacuously satisfiable -- 'no function named
+    acknowledge_all' is worthless if signing doesn't actually require
+    acknowledgement in practice).
+
+    1. No bulk-acknowledge shortcut exists (see the comment in
+       catalog_review.py above SignDecision -- deliberate friction).
+    2. The gate that friction protects is actually enforced: with entries
+       still unacknowledged, can_sign() must refuse, not just "some GUI
+       checkbox happens to be unticked".
+    """
+    names = [n for n in dir(r) if "acknowledge" in n.lower()]
+    assert "acknowledge_all" not in names
+    assert "acknowledge_all_entries" not in names
+
+    session = r.start_review("sha1", _catalog(), _catalog(_entry(id="a"), _entry(id="b")))
+    r.acknowledge_entry(session, "a")  # only one of two -- not a bulk call
+    decision = r.can_sign(session, "sha1")
+    assert decision.ok is False
+    assert "acknowledged" in decision.reason.lower()
+
+
+# --------------------------------------------------------------------------- #
+# can_sign: TOCTOU blob pinning + acknowledge gating combined
+# --------------------------------------------------------------------------- #
+def test_can_sign_false_when_not_all_acknowledged():
+    session = r.start_review("sha1", _catalog(), _catalog(_entry()))
+    decision = r.can_sign(session, "sha1")
+    assert decision.ok is False
+    assert "acknowledged" in decision.reason
+
+
+def test_can_sign_true_when_acknowledged_and_blob_matches():
+    session = r.start_review("sha1", _catalog(), _catalog(_entry()))
+    r.acknowledge_entry(session, "filesystem")
+    decision = r.can_sign(session, "sha1")
+    assert decision.ok is True
+    assert decision.reason is None
+
+
+def test_can_sign_refuses_on_blob_mismatch_even_if_acknowledged():
+    """The core TOCTOU fix: acknowledging everything is not enough if the
+    bytes on the remote changed underneath the review."""
+    session = r.start_review("sha1", _catalog(), _catalog(_entry()))
+    r.acknowledge_entry(session, "filesystem")
+    decision = r.can_sign(session, "sha2-a-new-commit-landed")
+    assert decision.ok is False
+    assert "changed" in decision.reason.lower() or "mismatch" in decision.reason.lower()
+
+
+def test_can_sign_blob_mismatch_takes_priority_message():
+    session = r.start_review("sha1", _catalog(), _catalog(_entry()))
+    decision = r.can_sign(session, "sha2")
+    assert decision.ok is False
+    assert "blob" in decision.reason.lower() or "changed" in decision.reason.lower()
+
+
+def test_can_sign_false_on_empty_changeset():
+    """The exact bug behind issue #68 finding 1: 'main' loaded against
+    itself diffs to [], and an empty changeset used to leave can_sign()
+    with nothing to refuse on (set() <= set() is vacuously True). Commit
+    b08cf21 signed 19 entries through precisely this path -- zero of them
+    were ever reviewed. An empty diff must mean 'nothing to sign', never
+    'sign unlocked'."""
+    same_catalog = _catalog(_entry())
+    session = r.start_review("sha1", same_catalog, same_catalog)
+    assert session.changes == []  # diff_catalogs(x, x) -> []
+    assert r.all_entries_acknowledged(session) is True  # vacuously -- this is the trap
+    decision = r.can_sign(session, "sha1")  # blob matches, "everything" acknowledged
+    assert decision.ok is False
+    assert "nothing to sign" in decision.reason.lower()
+
+
+def test_can_sign_false_with_outstanding_blocking_risk_even_if_acknowledged():
+    """can_sign() must itself refuse a blocking risk finding -- today a
+    blocking finding only disables the GUI checkbox, so the pure gate must
+    not simply trust that the caller never acknowledged a blocking entry.
+    Acknowledge it directly here (bypassing any GUI checkbox-disable logic
+    entirely) to prove the gate catches it independently of the GUI."""
+    session = r.start_review(
+        "sha1",
+        _catalog(),
+        _catalog(_entry(config={"command": "bash", "args": ["-c", "evil"]})),
+    )
+    r.acknowledge_entry(session, "filesystem")
+    assert r.all_entries_acknowledged(session) is True
+    decision = r.can_sign(session, "sha1")
+    assert decision.ok is False
+    assert "blocking" in decision.reason.lower()
+
+
+# --------------------------------------------------------------------------- #
+# sign_precondition: the ref-resolution seam that used to hardcode "main"
+# --------------------------------------------------------------------------- #
+def test_sign_precondition_resolves_against_loaded_ref_not_hardcoded_main():
+    """The regression test for issue #68 finding 1's first bug:
+    ReviewWindow._on_sign used to hardcode fetch_ref(repo, "main") as the
+    TOCTOU comparison ref. For a PR review, _on_load pins the PR HEAD's
+    blob SHA, so comparing against main's SHA differs by definition and
+    Sign could never fire on the PR path.
+
+    The fake resolver below returns a DIFFERENT (deliberately wrong) SHA for
+    "main" than for the PR ref that was actually loaded. If
+    sign_precondition ever resolves against "main" instead of
+    session.loaded_ref, this test fails -- both via the recorded `calls`
+    list and via decision.ok flipping to False.
+    """
+    pr_ref = "refs/pull/42/head"
+    session = r.start_review("pr-blob-sha", _catalog(), _catalog(_entry()), loaded_ref=pr_ref)
+    r.acknowledge_entry(session, "filesystem")
+
+    calls: list[str] = []
+
+    def fake_resolver(ref: str) -> str:
+        calls.append(ref)
+        return {"main": "main-blob-sha-WRONG", pr_ref: "pr-blob-sha"}[ref]
+
+    decision = r.sign_precondition(session, fake_resolver)
+    assert calls == [pr_ref]  # never asked the resolver for "main"
+    assert decision.ok is True
+    assert decision.reason is None
+
+
+def test_sign_precondition_refuses_when_loaded_ref_blob_moved():
+    """Same seam, the negative case: if the loaded ref's blob SHA has moved
+    since review began (a new commit landed on the reviewed PR/branch), the
+    resolver reflects that and sign_precondition must refuse -- proving this
+    isn't just a hardcoded pass-through."""
+    pr_ref = "refs/pull/42/head"
+    session = r.start_review("pr-blob-sha", _catalog(), _catalog(_entry()), loaded_ref=pr_ref)
+    r.acknowledge_entry(session, "filesystem")
+
+    def fake_resolver(_ref: str) -> str:
+        return "pr-blob-sha-AFTER-A-NEW-PUSH"
+
+    decision = r.sign_precondition(session, fake_resolver)
+    assert decision.ok is False
+    assert "mismatch" in decision.reason.lower() or "changed" in decision.reason.lower()
+
+
+def test_sign_precondition_defaults_to_main_when_loaded_ref_unset():
+    """start_review()'s loaded_ref defaults to 'main' for source=main
+    reviews (and backward-compat with callers that don't pass it)."""
+    session = r.start_review("sha1", _catalog(), _catalog(_entry()))
+    assert session.loaded_ref == "main"
+    r.acknowledge_entry(session, "filesystem")
+
+    def fake_resolver(ref: str) -> str:
+        assert ref == "main"
+        return "sha1"
+
+    decision = r.sign_precondition(session, fake_resolver)
+    assert decision.ok is True
+
+
+# --------------------------------------------------------------------------- #
+# find_last_signed_catalog_raw: what source=main diffs against
+# --------------------------------------------------------------------------- #
+def test_find_last_signed_catalog_raw_returns_matching_candidate():
+    """Simulates walking catalog.json's git history: the CURRENT signature
+    covers an OLDER version of the bytes (a later commit changed
+    catalog.json without re-signing -- the exact bypass that produced
+    commit b08cf21). The first candidate that verifies against that
+    signature is 'the last catalog a maintainer actually signed'."""
+    seed, pubkey = r.generate_keypair()
+    old_raw = b'{"schema":1,"version":1,"servers":[]}'
+    new_raw = b'{"schema":1,"version":2,"servers":[]}'
+    sig = r.sign_catalog_bytes(old_raw, seed)  # signature covers the OLD bytes
+    found = r.find_last_signed_catalog_raw([new_raw, old_raw], sig, [pubkey])
+    assert found == old_raw
+
+
+def test_find_last_signed_catalog_raw_none_when_nothing_verifies():
+    seed, _pubkey = r.generate_keypair()
+    _other_seed, other_pubkey = r.generate_keypair()
+    raw = b'{"schema":1,"version":1,"servers":[]}'
+    sig = r.sign_catalog_bytes(raw, seed)
+    # Check against a pubkey list that does NOT include the signer's key.
+    assert r.find_last_signed_catalog_raw([raw], sig, [other_pubkey]) is None
+
+
+# --------------------------------------------------------------------------- #
+# catalog_signing_message: domain separation must match bcc_core exactly
+# --------------------------------------------------------------------------- #
+def test_signing_message_uses_bcc_core_domain_prefix():
+    raw = b'{"schema":1}'
+    msg = r.catalog_signing_message(raw)
+    assert msg == c._CATALOG_SIG_DOMAIN + raw
+    assert msg.startswith(b"bcc-catalog-v1|")
+
+
+def test_sign_then_verify_round_trips_with_bcc_core():
+    """End-to-end: a signature produced by the Console's sign_catalog_bytes
+    must verify with bcc_core.verify_catalog_signature -- proves the two
+    modules can never drift on the domain-separation prefix."""
+    seed, pubkey = r.generate_keypair()
+    raw = b'{"schema":1,"version":2,"servers":[]}'
+    sig = r.sign_catalog_bytes(raw, seed)
+    assert c.verify_catalog_signature(raw, sig, [pubkey]) is True
+
+
+def test_sign_tampered_bytes_fails_verify():
+    seed, pubkey = r.generate_keypair()
+    raw = b'{"schema":1,"version":2,"servers":[]}'
+    sig = r.sign_catalog_bytes(raw, seed)
+    tampered = raw[:-1] + b"0"
+    assert c.verify_catalog_signature(tampered, sig, [pubkey]) is False
+
+
+# --------------------------------------------------------------------------- #
+# Key encryption at rest
+# --------------------------------------------------------------------------- #
+def test_encrypt_decrypt_round_trip():
+    seed, _pub = r.generate_keypair()
+    blob = r.encrypt_private_key(seed, "correct horse battery staple")
+    decrypted = r.decrypt_private_key(blob, "correct horse battery staple")
+    assert decrypted == seed
+
+
+def test_decrypt_wrong_passphrase_raises():
+    seed, _pub = r.generate_keypair()
+    blob = r.encrypt_private_key(seed, "right passphrase")
+    with pytest.raises(ValueError):
+        r.decrypt_private_key(blob, "wrong passphrase")
+
+
+def test_decrypt_corrupted_blob_raises():
+    seed, _pub = r.generate_keypair()
+    blob = r.encrypt_private_key(seed, "pass")
+    corrupted = blob[:-1] + bytes([blob[-1] ^ 0xFF])
+    with pytest.raises(ValueError):
+        r.decrypt_private_key(corrupted, "pass")
+
+
+def test_encrypt_private_key_rejects_wrong_length_seed():
+    with pytest.raises(ValueError):
+        r.encrypt_private_key(b"too-short", "pass")
+
+
+def test_encrypt_private_key_rejects_empty_passphrase():
+    seed, _pub = r.generate_keypair()
+    with pytest.raises(ValueError):
+        r.encrypt_private_key(seed, "")
+
+
+def test_encrypted_blob_never_contains_seed_plaintext():
+    seed, _pub = r.generate_keypair()
+    blob = r.encrypt_private_key(seed, "some passphrase")
+    assert seed not in blob
+
+
+def test_generate_keypair_produces_valid_ed25519_pair():
+    seed, pubkey = r.generate_keypair()
+    assert len(seed) == 32
+    assert len(pubkey) == 32
+    raw = b"test payload"
+    sig = r.sign_catalog_bytes(raw, seed)
+    assert c.verify_catalog_signature(raw, sig, [pubkey]) is True
+
+
+# --------------------------------------------------------------------------- #
+# Registry lookup / near-neighbour edit distance
+# --------------------------------------------------------------------------- #
+def test_edit_distance_identical():
+    assert r.edit_distance("abc", "abc") == 0
+
+
+def test_edit_distance_one_substitution():
+    assert r.edit_distance("firecrawl-mcp", "f1recrawl-mcp") == 1
+
+
+def test_near_neighbor_ids_finds_close_match():
+    others = ["firecrawl-mcp", "unrelated-server", "totally-different"]
+    neighbors = r.near_neighbor_ids("firecrawl-mcp2", others, max_distance=2)
+    assert "firecrawl-mcp" in neighbors
+
+
+def test_near_neighbor_ids_excludes_self():
+    others = ["filesystem", "other"]
+    assert "filesystem" not in r.near_neighbor_ids("filesystem", others)
+
+
+def test_near_neighbor_ids_excludes_far_matches():
+    others = ["completely-unrelated-name"]
+    assert r.near_neighbor_ids("filesystem", others, max_distance=2) == []
+
+
+def test_extract_package_refs_npm():
+    entry = _entry(config={"command": "npx", "args": ["-y", "@scope/pkg@1.2.3"]})
+    refs = r.extract_package_refs(entry)
+    assert len(refs) == 1
+    assert refs[0].ecosystem == "npm"
+    assert refs[0].name == "@scope/pkg"
+    assert refs[0].version == "1.2.3"
+
+
+def test_extract_package_refs_uvx():
+    entry = _entry(config={"command": "uvx", "args": ["some-pypi-pkg==1.0.0"]})
+    refs = r.extract_package_refs(entry)
+    assert len(refs) == 1
+    assert refs[0].ecosystem == "pypi"
+    assert refs[0].name == "some-pypi-pkg"
+    assert refs[0].version == "1.0.0"
+
+
+def test_extract_package_refs_link_only_entry_returns_empty():
+    entry = {"id": "slack", "setup": "link-only", "docs_url": "https://example.com"}
+    assert r.extract_package_refs(entry) == []
+
+
+def test_lookup_registry_info_fails_soft_on_none():
+    ref = r.PackageRef("x", "npm", "somepkg", "1.0.0")
+    info = r.lookup_registry_info(ref, lambda _ref: None, [])
+    assert info.available is False
+
+
+def test_lookup_registry_info_fails_soft_on_exception():
+    def boom(_ref):
+        raise RuntimeError("network down")
+
+    ref = r.PackageRef("x", "npm", "somepkg", "1.0.0")
+    info = r.lookup_registry_info(ref, boom, [])
+    assert info.available is False
+
+
+def test_lookup_registry_info_populates_fields_when_available():
+    ref = r.PackageRef("x", "npm", "somepkg", "1.0.0")
+
+    def fetcher(_ref):
+        return {
+            "publisher": "hello_sideguide",
+            "age_days": 30,
+            "last_release": "2026-01-01",
+            "downloads": 500,
+        }
+
+    info = r.lookup_registry_info(ref, fetcher, [])
+    assert info.available is True
+    assert info.publisher == "hello_sideguide"
+    assert info.downloads == 500
+
+
+def test_lookup_registry_info_includes_near_neighbors():
+    ref = r.PackageRef("x", "npm", "firecrawl-mcp2", "1.0.0")
+    info = r.lookup_registry_info(ref, lambda _ref: None, ["firecrawl-mcp"])
+    assert "firecrawl-mcp" in info.near_neighbor_ids
+
+
+# --------------------------------------------------------------------------- #
+# contains_non_ascii
+# --------------------------------------------------------------------------- #
+def test_contains_non_ascii_true():
+    assert r.contains_non_ascii("pаckage") is True
+
+
+def test_contains_non_ascii_false():
+    assert r.contains_non_ascii("package") is False

tests/test_core.py

@@ -1,6 +1,7 @@
 """Pytest port of the original test_core.py script (same 23 behaviours, now
 proper test functions with tmp_path/monkeypatch fixtures)."""
 
+import dataclasses
 import json
 import os
 import re
@@ -1690,8 +1691,12 @@ def _minimal_catalog(version: int = 1) -> dict:
                 "official": True,
                 "setup": "basic",
                 "config": {
+                    # Pinned on purpose (issue #68 finding 3): an earlier
+                    # version of this fixture used an unpinned package and
+                    # asserted it validated clean, which enshrined the bug
+                    # instead of catching it.
                     "command": "npx",
-                    "args": ["-y", "widget-mcp"],
+                    "args": ["-y", "widget-mcp@1.0.0"],
                 },
                 "placeholders": {},
                 "env_required": {},
@@ -1952,6 +1957,213 @@ def test_validate_catalog_rejects_duplicate_ids():
     assert any("duplicate id" in p for p in problems)
 
 
+# --- validate_catalog: config.env (issue #68 finding 2) -------------------- #
+def test_validate_catalog_rejects_each_denied_env_key():
+    for key in sorted(c.CATALOG_DENIED_ENV_KEYS):
+        data = _catalog_with(
+            {"config": {"command": "npx", "args": ["-y", "widget-mcp@1.0.0"], "env": {key: ""}}}
+        )
+        problems = c.validate_catalog(data)
+        assert any("deny-list" in p for p in problems), (key, problems)
+
+
+def test_validate_catalog_rejects_denied_env_key_case_insensitively():
+    data = _catalog_with(
+        {
+            "config": {
+                "command": "npx",
+                "args": ["-y", "widget-mcp@1.0.0"],
+                "env": {"node_options": ""},
+            }
+        }
+    )
+    problems = c.validate_catalog(data)
+    assert any("deny-list" in p for p in problems)
+
+
+def test_validate_catalog_rejects_nonempty_nonplaceholder_env_value():
+    data = _catalog_with(
+        {
+            "config": {
+                "command": "npx",
+                "args": ["-y", "widget-mcp@1.0.0"],
+                "env": {"FOO_URL": "https://example.com"},
+            }
+        }
+    )
+    problems = c.validate_catalog(data)
+    assert any("empty string or a single <PLACEHOLDER>" in p for p in problems)
+
+
+def test_validate_catalog_accepts_empty_and_placeholder_env_values():
+    data = _catalog_with(
+        {
+            "config": {
+                "command": "npx",
+                "args": ["-y", "widget-mcp@1.0.0"],
+                "env": {"FOO": "", "BAR_URL": "<BAR_URL>"},
+            }
+        }
+    )
+    assert c.validate_catalog(data) == []
+
+
+def test_validate_catalog_rejects_non_ascii_env_key():
+    data = _catalog_with(
+        {
+            "config": {
+                "command": "npx",
+                "args": ["-y", "widget-mcp@1.0.0"],
+                "env": {"FÖO": ""},
+            }
+        }
+    )
+    problems = c.validate_catalog(data)
+    assert any("config.env key" in p and "ASCII" in p for p in problems)
+
+
+def test_validate_catalog_rejects_non_ascii_env_value():
+    data = _catalog_with(
+        {
+            "config": {
+                "command": "npx",
+                "args": ["-y", "widget-mcp@1.0.0"],
+                "env": {"FOO": "<Bäd>"},
+            }
+        }
+    )
+    problems = c.validate_catalog(data)
+    assert any("config.env value" in p and "ASCII" in p for p in problems)
+
+
+def test_validate_catalog_rejects_secret_looking_env_value():
+    data = _catalog_with(
+        {
+            "config": {
+                "command": "npx",
+                "args": ["-y", "widget-mcp@1.0.0"],
+                "env": {"SOME_TOKEN": "ghp_abcdef1234567890"},
+            }
+        }
+    )
+    problems = c.validate_catalog(data)
+    assert any("real secret value" in p for p in problems)
+
+
+def test_validate_catalog_rejects_node_options_env_walking_past_allowlist():
+    # The exact reproduction from issue #68 finding 2: an allowlisted
+    # `npx` command carrying NODE_OPTIONS in env, which previously passed
+    # validation and would have flowed straight into the executed
+    # subprocess via catalog_entry_to_paste_json().
+    data = _catalog_with(
+        {
+            "config": {
+                "command": "npx",
+                "args": ["-y", "widget-mcp@1.0.0"],
+                "env": {"NODE_OPTIONS": "--require /tmp/payload.js"},
+            }
+        }
+    )
+    problems = c.validate_catalog(data)
+    assert problems != []
+
+
+# --- validate_catalog: version pinning (issue #68 finding 3) --------------- #
+def test_validate_catalog_rejects_unpinned_npx_package():
+    data = _catalog_with({"config": {"command": "npx", "args": ["-y", "widget-mcp"]}})
+    problems = c.validate_catalog(data)
+    assert any("not version-pinned" in p for p in problems)
+
+
+def test_validate_catalog_rejects_unpinned_scoped_npx_package():
+    data = _catalog_with({"config": {"command": "npx", "args": ["-y", "@scope/pkg"]}})
+    problems = c.validate_catalog(data)
+    assert any("not version-pinned" in p for p in problems)
+
+
+def test_validate_catalog_accepts_pinned_scoped_npx_package():
+    data = _catalog_with({"config": {"command": "npx", "args": ["-y", "@scope/pkg@1.2.3"]}})
+    assert c.validate_catalog(data) == []
+
+
+def test_validate_catalog_rejects_unpinned_uvx_package():
+    data = _catalog_with({"config": {"command": "uvx", "args": ["some-tool"]}})
+    problems = c.validate_catalog(data)
+    assert any("not version-pinned" in p for p in problems)
+
+
+def test_validate_catalog_accepts_uvx_at_version_pin():
+    data = _catalog_with({"config": {"command": "uvx", "args": ["some-tool@1.0.0"]}})
+    assert c.validate_catalog(data) == []
+
+
+def test_validate_catalog_accepts_uvx_double_equals_pin():
+    data = _catalog_with({"config": {"command": "uvx", "args": ["some-tool==1.0.0"]}})
+    assert c.validate_catalog(data) == []
+
+
+def test_validate_catalog_rejects_docker_latest_tag():
+    data = _catalog_with({"config": {"command": "docker", "args": ["run", "some/image:latest"]}})
+    problems = c.validate_catalog(data)
+    assert any("'latest'" in p for p in problems)
+
+
+def test_validate_catalog_rejects_docker_untagged_image():
+    data = _catalog_with({"config": {"command": "docker", "args": ["run", "some/image"]}})
+    problems = c.validate_catalog(data)
+    assert any("no explicit tag" in p for p in problems)
+
+
+def test_validate_catalog_accepts_pinned_docker_image_with_flags():
+    data = _catalog_with(
+        {
+            "config": {
+                "command": "docker",
+                "args": ["run", "-i", "--rm", "-e", "SOME_TOKEN", "some/image:1.2.3"],
+            }
+        }
+    )
+    assert c.validate_catalog(data) == []
+
+
+def test_validate_catalog_does_not_pin_check_placeholders_flags_or_subcommand():
+    # A pinned uvx spec followed by flags and a <PLACEHOLDER> positional
+    # must not itself get mistaken for an unpinned package.
+    data = _catalog_with(
+        {
+            "config": {
+                "command": "uvx",
+                "args": ["mcp-server-git@2026.7.10", "--repository", "<REPO_PATH>"],
+            }
+        }
+    )
+    assert c.validate_catalog(data) == []
+
+
+# --- validate_catalog: id constraint (issue #68 finding 7) ----------------- #
+def test_validate_catalog_rejects_id_with_markup():
+    data = _catalog_with({"id": "<b>Verified</b>"})
+    problems = c.validate_catalog(data)
+    assert any("must match" in p for p in problems)
+
+
+def test_validate_catalog_rejects_id_with_uppercase():
+    data = _catalog_with({"id": "Widget"})
+    problems = c.validate_catalog(data)
+    assert any("must match" in p for p in problems)
+
+
+def test_validate_catalog_rejects_id_starting_with_dash():
+    data = _catalog_with({"id": "-widget"})
+    problems = c.validate_catalog(data)
+    assert any("must match" in p for p in problems)
+
+
+def test_validate_catalog_accepts_valid_slug_id():
+    data = _catalog_with({"id": "widget-2.thing-ok"})
+    assert c.validate_catalog(data) == []
+
+
 # --- resolve_catalog -------------------------------------------------------- #
 def test_resolve_catalog_nothing_available_returns_empty_dict():
     assert c.resolve_catalog(None, None, None) == {}
@@ -2003,13 +2215,93 @@ def test_resolve_catalog_rejects_absurd_version_jump(monkeypatch):
     pub = priv.public_key().public_bytes_raw()
     monkeypatch.setattr(c, "CATALOG_PUBKEYS", [pub])
 
-    cached = _signed(_minimal_catalog(version=5), priv)
+    # The freeze attempt goes FIRST (as `cached`), the legitimate catalog
+    # SECOND (as `remote`) -- on purpose. Putting the good catalog first
+    # (as an earlier version of this test did) never exercises the
+    # vulnerable path: the old implementation only guarded a candidate
+    # against "the best accepted so far," so whichever candidate was
+    # evaluated FIRST got in unconditionally, uncapped. Ordering the freeze
+    # attempt first is what actually proves the cap holds regardless of
+    # evaluation order.
     freeze_attempt = _signed(_minimal_catalog(version=999999), priv)
+    good = _signed(_minimal_catalog(version=5), priv)
 
-    result = c.resolve_catalog(None, cached, freeze_attempt)
+    result = c.resolve_catalog(None, freeze_attempt, good)
     assert c.catalog_version(result) == 5
 
 
+def test_resolve_catalog_caps_first_and_only_candidate(monkeypatch):
+    # issue #68 finding 6: with no bundled catalog to anchor against, a
+    # signed catalog claiming an absurd version must still be capped even
+    # when it is the ONLY candidate resolve_catalog() ever sees -- there is
+    # no "best so far" for it to be compared against, so the cap has to
+    # apply unconditionally, not "once something else has already landed."
+    priv = Ed25519PrivateKey.generate()
+    pub = priv.public_key().public_bytes_raw()
+    monkeypatch.setattr(c, "CATALOG_PUBKEYS", [pub])
+
+    freeze_attempt = _signed(_minimal_catalog(version=999999999), priv)
+
+    result = c.resolve_catalog(None, None, freeze_attempt)
+    assert result == {}
+
+
+def test_resolve_catalog_anchors_cap_to_bundled_not_a_chained_best(monkeypatch):
+    # Anti-freeze must be measured against the BUNDLED version specifically,
+    # not against "whatever the best-so-far happens to be after each
+    # candidate is accepted" -- a chained anchor lets each accepted
+    # candidate ratchet the allowed ceiling upward, so a legitimate
+    # moderate bump (cached) plus a second, much larger jump (remote) can
+    # each individually look "within _CATALOG_MAX_VERSION_JUMP of the
+    # previous one" while remote is nowhere near bundled's version.
+    priv = Ed25519PrivateKey.generate()
+    pub = priv.public_key().public_bytes_raw()
+    monkeypatch.setattr(c, "CATALOG_PUBKEYS", [pub])
+
+    bundled = _signed(_minimal_catalog(version=2), priv)
+    cached = _signed(_minimal_catalog(version=1000), priv)  # within 1000 of bundled
+    remote = _signed(_minimal_catalog(version=1900), priv)  # within 1000 of cached,
+    #                                                          NOT of bundled
+
+    result = c.resolve_catalog(bundled, cached, remote)
+    assert c.catalog_version(result) == 1000
+
+
+def test_resolve_catalog_prefers_bundled_on_version_tie(monkeypatch):
+    # issue #68 finding 6: on a tie the LAST candidate evaluated used to
+    # win, so remote silently beat bundled at equal version. Bundled --
+    # the copy frozen into the binary -- must win ties.
+    priv = Ed25519PrivateKey.generate()
+    pub = priv.public_key().public_bytes_raw()
+    monkeypatch.setattr(c, "CATALOG_PUBKEYS", [pub])
+
+    bundled_data = _minimal_catalog(version=5)
+    bundled = _signed(bundled_data, priv)
+
+    remote_data = _minimal_catalog(version=5)
+    remote_data["servers"][0]["display"] = "Remote Impostor"
+    remote = _signed(remote_data, priv)
+
+    result = c.resolve_catalog(bundled, None, remote)
+    assert c.catalog_version(result) == 5
+    assert result["servers"][0]["display"] == "Widget"
+
+
+def test_resolve_catalog_floor_rejects_below_persisted_version(monkeypatch):
+    # `floor` is a pure parameter: the caller (eventually the GUI, from
+    # persisted storage) can pass a previously-accepted version, and
+    # nothing below it may be accepted even with no bundled catalog to
+    # anchor against.
+    priv = Ed25519PrivateKey.generate()
+    pub = priv.public_key().public_bytes_raw()
+    monkeypatch.setattr(c, "CATALOG_PUBKEYS", [pub])
+
+    stale = _signed(_minimal_catalog(version=3), priv)
+
+    result = c.resolve_catalog(None, None, stale, floor=10)
+    assert result == {}
+
+
 def test_resolve_catalog_malformed_candidate_does_not_raise(monkeypatch):
     priv = Ed25519PrivateKey.generate()
     pub = priv.public_key().public_bytes_raw()
@@ -2039,15 +2331,32 @@ def test_resolve_catalog_invalid_but_signed_candidate_is_skipped(monkeypatch):
 def test_catalog_entry_to_paste_json_basic_shape():
     entry = _minimal_catalog()["servers"][0]
     result = c.catalog_entry_to_paste_json(entry)
-    assert result == {"widget": {"command": "npx", "args": ["-y", "widget-mcp"]}}
+    assert result == {"widget": {"command": "npx", "args": ["-y", "widget-mcp@1.0.0"]}}
 
 
 def test_catalog_entry_to_paste_json_includes_env_when_present():
+    # NOTE: catalog_entry_to_paste_json() is a pure shape-converter for an
+    # entry that has ALREADY passed validate_catalog() -- it is correct for
+    # it to carry env through verbatim. The bug (issue #68 finding 2) was
+    # never in this function; it was that validate_catalog() let entries
+    # with dangerous/non-placeholder env values reach this function in the
+    # first place. This test now proves that boundary explicitly: a
+    # validation-legal env value (a <PLACEHOLDER> token) survives the
+    # conversion, and a value validate_catalog() would have rejected is
+    # confirmed rejected before it ever gets here.
     entry = _minimal_catalog()["servers"][0]
     entry["config"]["env"] = {"GRAFANA_URL": "<GRAFANA_URL>"}
     result = c.catalog_entry_to_paste_json(entry)
     assert result["widget"]["env"] == {"GRAFANA_URL": "<GRAFANA_URL>"}
 
+    catalog = _minimal_catalog()
+    catalog["servers"][0]["config"]["env"] = {"GRAFANA_URL": "<GRAFANA_URL>"}
+    assert c.validate_catalog(catalog) == []
+
+    malicious = _minimal_catalog()
+    malicious["servers"][0]["config"]["env"] = {"NODE_OPTIONS": "--require /tmp/payload.js"}
+    assert c.validate_catalog(malicious) != []
+
 
 def test_config_has_unfilled_placeholders_true_for_token():
     cfg = {"command": "npx", "args": ["-y", "server", "<ALLOWED_DIR>"]}
@@ -2062,3 +2371,312 @@ def test_config_has_unfilled_placeholders_false_after_fill():
 def test_config_has_unfilled_placeholders_checks_env_too():
     cfg = {"command": "uvx", "args": ["mcp-grafana"], "env": {"GRAFANA_URL": "<GRAFANA_URL>"}}
     assert c.config_has_unfilled_placeholders(cfg) is True
+
+
+# --------------------------------------------------------------------------- #
+# #72 -- a server value that isn't a JSON object must not take the load down
+# --------------------------------------------------------------------------- #
+@pytest.mark.parametrize("bad", ["not-a-dict", 123, ["a", "b"], None, True, 1.5])
+def test_extract_servers_survives_non_dict_server_value(bad):
+    entries = c.extract_servers({"mcpServers": {"foo": bad}})
+    assert len(entries) == 1
+    assert entries[0].name == "foo"
+    assert entries[0].data == {}
+    assert entries[0].malformed is True
+    assert entries[0].raw == bad
+
+
+def test_extract_servers_marks_only_the_bad_entry():
+    cfg = {"mcpServers": {"good": {"command": "npx"}, "bad": "oops"}}
+    by_name = {e.name: e for e in c.extract_servers(cfg)}
+    assert by_name["good"].malformed is False
+    assert by_name["good"].data == {"command": "npx"}
+    assert by_name["bad"].malformed is True
+
+
+def test_extract_servers_handles_malformed_disabled_entry():
+    entries = c.extract_servers({c.DISABLED_KEY: {"parked": ["nope"]}})
+    assert entries[0].enabled is False
+    assert entries[0].malformed is True
+
+
+def test_malformed_entry_round_trips_through_save_unchanged():
+    """The cardinal rule: never silently delete what the user had on disk."""
+    cfg = {"mcpServers": {"good": {"command": "npx"}, "bad": "oops"}}
+    servers = c.extract_servers(cfg)
+    out = c.apply_servers(dict(cfg), servers)
+    assert out["mcpServers"]["bad"] == "oops"
+    assert out["mcpServers"]["good"] == {"command": "npx"}
+
+
+def test_editing_a_malformed_entry_retires_the_raw_value():
+    entry = c.extract_servers({"mcpServers": {"bad": "oops"}})[0]
+    entry.set_data({"command": "npx"})
+    assert entry.malformed is False
+    assert entry.config_value() == {"command": "npx"}
+    assert c.apply_servers({}, [entry])["mcpServers"]["bad"] == {"command": "npx"}
+
+
+def test_lint_reports_the_malformed_entry_by_name():
+    servers = c.extract_servers({"mcpServers": {"bad": "oops"}})
+    warnings = c.lint_servers(servers)
+    assert len(warnings) == 1
+    assert "'bad'" in warnings[0]
+    assert "not an object" in warnings[0]
+    assert "str" in warnings[0]
+
+
+def test_lint_still_reports_normal_warnings_alongside_malformed():
+    cfg = {"mcpServers": {"bad": "oops", "sloppy": {"command": "npx", "args": "one two"}}}
+    warnings = c.lint_servers(c.extract_servers(cfg))
+    assert any("not an object" in w for w in warnings)
+    assert any("'args' should be a list" in w for w in warnings)
+
+
+# --------------------------------------------------------------------------- #
+# #73 -- the stale-file merge must not discard BCC-authored keys
+# --------------------------------------------------------------------------- #
+def test_carry_owned_keys_moves_sets_onto_the_reloaded_config():
+    local = {"mcpServers": {}, c.SETS_KEY: {"work": ["a", "b"]}}
+    fresh = {"mcpServers": {"external": {"command": "npx"}}}
+    contested = c.carry_owned_keys(local, fresh)
+    assert contested == []
+    assert fresh[c.SETS_KEY] == {"work": ["a", "b"]}
+    assert fresh["mcpServers"] == {"external": {"command": "npx"}}
+
+
+def test_carry_owned_keys_reports_a_genuine_conflict():
+    local = {c.SETS_KEY: {"work": ["a"]}}
+    fresh = {c.SETS_KEY: {"work": ["a", "b"]}}
+    assert c.carry_owned_keys(local, fresh) == [c.SETS_KEY]
+    assert fresh[c.SETS_KEY] == {"work": ["a"]}  # local wins: BCC owns the key
+
+
+def test_carry_owned_keys_is_quiet_when_both_sides_agree():
+    local = {c.SETS_KEY: {"work": ["a"]}}
+    fresh = {c.SETS_KEY: {"work": ["a"]}}
+    assert c.carry_owned_keys(local, fresh) == []
+
+
+def test_carry_owned_keys_leaves_disk_alone_when_absent_locally():
+    """Can't distinguish 'deleted my last set' from 'never had sets'; keep theirs."""
+    fresh = {c.SETS_KEY: {"remote": ["a"]}}
+    assert c.carry_owned_keys({}, fresh) == []
+    assert fresh[c.SETS_KEY] == {"remote": ["a"]}
+
+
+def test_carry_owned_keys_deep_copies_so_later_edits_do_not_leak():
+    local = {c.SETS_KEY: {"work": ["a"]}}
+    fresh = {}
+    c.carry_owned_keys(local, fresh)
+    local[c.SETS_KEY]["work"].append("b")
+    assert fresh[c.SETS_KEY] == {"work": ["a"]}
+
+
+def test_merge_flow_preserves_sets_and_external_servers(tmp_path):
+    """End-to-end shape of the Merge & save path that lost sets in #73."""
+    path = tmp_path / "claude.json"
+    path.write_text(json.dumps({"mcpServers": {"old": {"command": "old"}}}))
+
+    # BCC loads, user saves a named set and edits servers in memory.
+    local = c.load_config(path)
+    servers = c.extract_servers(local)
+    c.save_server_set(local, "work", servers)
+
+    # Something else rewrites the file underneath us.
+    path.write_text(json.dumps({"mcpServers": {"external": {"command": "new"}}, "other": 1}))
+
+    # Merge & save: reload disk, carry BCC keys, re-apply the user's servers.
+    fresh = c.load_config(path)
+    c.carry_owned_keys(local, fresh)
+    c.apply_servers(fresh, servers)
+    c.write_config(path, fresh)
+
+    saved = c.load_config(path)
+    assert saved[c.SETS_KEY] == {"work": ["old"]}  # the set survived
+    assert saved["other"] == 1  # unrelated external key preserved
+    assert "old" in saved["mcpServers"]  # user's servers re-applied
+
+
+def test_null_server_value_is_malformed_not_mistaken_for_absent():
+    """`{"mcpServers": {"foo": null}}` is legal JSON and a real malformed case,
+    so None must not double as the 'nothing here' sentinel."""
+    entry = c.extract_servers({"mcpServers": {"foo": None}})[0]
+    assert entry.malformed is True
+    assert entry.raw is None
+    assert c.apply_servers({}, [entry])["mcpServers"]["foo"] is None
+
+
+def test_a_normal_entry_is_not_malformed():
+    entry = c.extract_servers({"mcpServers": {"foo": {"command": "npx"}}})[0]
+    assert entry.malformed is False
+    assert entry.raw is c.NO_RAW
+
+
+# --------------------------------------------------------------------------- #
+# #75 -- theming
+# --------------------------------------------------------------------------- #
+@pytest.mark.parametrize(
+    "setting,system_dark,expected",
+    [
+        (c.THEME_DARK, False, "dark"),
+        (c.THEME_DARK, True, "dark"),
+        (c.THEME_LIGHT, False, "light"),
+        (c.THEME_LIGHT, True, "light"),
+        (c.THEME_SYSTEM, True, "dark"),
+        (c.THEME_SYSTEM, False, "light"),
+    ],
+)
+def test_resolve_theme_covers_every_setting_and_appearance(setting, system_dark, expected):
+    assert c.resolve_theme(setting, system_dark) == expected
+
+
+@pytest.mark.parametrize("junk", ["", "solarized", None, "DARK", 3])
+def test_resolve_theme_falls_back_to_following_the_system(junk):
+    """A hand-edited or future QSettings value should follow the desktop,
+    not pin a fixed theme."""
+    assert c.resolve_theme(junk, True) == "dark"
+    assert c.resolve_theme(junk, False) == "light"
+
+
+def test_palette_for_known_names():
+    assert c.palette_for("dark") is c.DARK_PALETTE
+    assert c.palette_for("light") is c.LIGHT_PALETTE
+
+
+def test_palette_for_unknown_name_falls_back_to_dark():
+    assert c.palette_for("chartreuse") is c.DARK_PALETTE
+
+
+def test_dark_palette_is_unchanged_from_the_shipped_look():
+    """v1.3.0 shipped these exact colours; adding a light theme must not
+    quietly restyle the dark one."""
+    p = c.DARK_PALETTE
+    assert (p.accent, p.bg, p.panel, p.panel_2) == ("#f97316", "#1b1d23", "#23262e", "#2b2f39")
+    assert (p.text, p.muted, p.border) == ("#e7e9ee", "#9aa0ad", "#3a3f4b")
+    assert (p.good, p.bad, p.warn, p.remote) == ("#4ade80", "#f87171", "#fbbf24", "#60a5fa")
+    assert (p.on_accent, p.disabled_bg, p.mono_bg) == ("#1a1205", "#202229", "#16181d")
+
+
+def test_both_palettes_define_every_slot():
+    """A missing slot should fail here rather than render a broken window."""
+    for pal in (c.DARK_PALETTE, c.LIGHT_PALETTE):
+        for f in dataclasses.fields(c.Palette):
+            value = getattr(pal, f.name)
+            assert value, f"{pal.name}.{f.name} is empty"
+            if f.name != "name":
+                assert re.fullmatch(r"#[0-9a-fA-F]{6}", value), f"{pal.name}.{f.name}={value!r}"
+
+
+@pytest.mark.parametrize("pal_name", ["dark", "light"])
+@pytest.mark.parametrize("slot", ["text", "muted", "good", "bad", "warn", "remote", "accent"])
+def test_palette_meets_contrast_on_panel(pal_name, slot):
+    """Every colour drawn as text/glyph must clear WCAG AA (4.5:1) against the
+    surface it sits on. The light palette's semantic colours are NOT the dark
+    ones lightened -- #4ade80 sits near 1.7:1 on white -- so this guards
+    against someone 'harmonising' them back toward the dark hues."""
+    pal = c.palette_for(pal_name)
+    assert c.contrast_ratio(getattr(pal, slot), pal.panel) >= 4.5
+
+
+@pytest.mark.parametrize("pal_name", ["dark", "light"])
+def test_on_accent_is_legible_against_the_accent_fill(pal_name):
+    """Primary buttons and selected rows draw on_accent on top of accent."""
+    pal = c.palette_for(pal_name)
+    assert c.contrast_ratio(pal.on_accent, pal.accent) >= 4.5
+
+
+def test_contrast_ratio_endpoints():
+    assert c.contrast_ratio("#000000", "#ffffff") == pytest.approx(21.0, abs=0.01)
+    assert c.contrast_ratio("#123456", "#123456") == pytest.approx(1.0, abs=0.001)
+    assert c.contrast_ratio("#ffffff", "#000000") == pytest.approx(21.0, abs=0.01)
+
+
+def test_relative_luminance_extremes():
+    assert c.relative_luminance("#000000") == pytest.approx(0.0)
+    assert c.relative_luminance("#ffffff") == pytest.approx(1.0)
+
+
+def test_stylesheet_builder_has_no_hardcoded_colours():
+    """Every colour in the QSS must come from the palette.
+
+    Three near-black literals used to be inlined here (#1a1205, #202229,
+    #16181d). Harmless with one theme; with two, they silently render dark
+    chrome on a light window. Reads the source rather than importing bcc,
+    which needs PySide6.
+    """
+    src = (Path(__file__).resolve().parent.parent / "bcc.py").read_text(encoding="utf-8")
+    start = src.index("def build_stylesheet")
+    body = src[start : src.index("def apply_palette")]
+    assert re.findall(r"#[0-9a-fA-F]{6}", body) == []
+
+
+def test_every_palette_slot_is_consumed():
+    """A slot added to Palette but never wired up is dead weight.
+
+    Checks for `p.<slot>` anywhere in bcc.py, which covers both the QSS and
+    apply_palette's global bindings -- not every slot belongs in the
+    stylesheet (`good` and `remote` feed the inline status dots via
+    STATUS_COLORS/HEALTH_COLORS, never the QSS). This won't catch a slot bound
+    to a global that nothing then uses; it does catch the common mistake of
+    extending the dataclass and forgetting to plumb it through.
+    """
+    src = (Path(__file__).resolve().parent.parent / "bcc.py").read_text(encoding="utf-8")
+    for f in dataclasses.fields(c.Palette):
+        if f.name == "name":
+            continue
+        assert f"p.{f.name}" in src, f"palette slot {f.name!r} is never consumed"
+
+
+# --------------------------------------------------------------------------- #
+# #78/#79 -- update notice: when to show it, and what it says
+# --------------------------------------------------------------------------- #
+def test_update_notice_when_a_newer_release_exists():
+    n = c.update_notice("1.2.0", {"version": "v1.3.0", "url": "https://example.test/rel"})
+    assert n is not None
+    assert n["version"] == "v1.3.0"
+    assert n["url"] == "https://example.test/rel"
+    assert "1.3.0" in n["text"] and "1.2.0" in n["text"]
+
+
+def test_update_notice_is_silent_when_current():
+    assert c.update_notice("1.3.0", {"version": "v1.3.0"}) is None
+    assert c.update_notice("1.4.0", {"version": "v1.3.0"}) is None
+
+
+@pytest.mark.parametrize("bad", [None, {}, {"version": ""}, {"version": None}, {"version": 3}, []])
+def test_update_notice_is_silent_on_a_failed_or_malformed_check(bad):
+    """fetch_latest_release returns None on any failure; a half-formed payload
+    must not produce a notice pointing at nothing."""
+    assert c.update_notice("1.0.0", bad) is None
+
+
+def test_update_notice_falls_back_to_the_releases_page_without_a_url():
+    n = c.update_notice("1.0.0", {"version": "v2.0.0"})
+    assert n["url"] == c.RELEASES_URL
+
+
+def test_update_notice_names_no_menu_path():
+    """The old status-line text said 'Help > About to view it', which is wrong
+    on macOS -- Qt moves the About action into the application menu (#79). The
+    notice carries its own action, so it must not describe a menu path."""
+    n = c.update_notice("1.0.0", {"version": "v2.0.0"})
+    lowered = n["text"].lower()
+    for phrase in ("help", "about", "menu", "▸", ">"):
+        assert phrase not in lowered, f"notice text should not reference {phrase!r}"
+
+
+def test_update_notice_handles_the_v_prefix_consistently():
+    assert c.update_notice("1.2.0", {"version": "1.3.0"}) is not None
+    assert c.update_notice("v1.2.0", {"version": "v1.3.0"}) is not None
+    assert c.update_notice("1.3.0", {"version": "v1.3.0"}) is None
+
+
+def test_update_notice_renders_both_versions_the_same_way():
+    """Tags carry a 'v' prefix, __version__ doesn't -- don't show both forms
+    in one sentence."""
+    n = c.update_notice("1.2.0", {"version": "v1.3.0"})
+    assert "v1.3.0" not in n["text"]
+    assert "1.3.0" in n["text"] and "1.2.0" in n["text"]
+    # the machine-readable field keeps the real tag
+    assert n["version"] == "v1.3.0"