Add legacy diff messages with inline and expanded viewing

[wesbillman]

Opened by Brain on Wes's behalf.

Summary

• Admit legacy kind-40008 diff messages through existing history/live/thread/exact/search/unread paths, preserving raw patch bytes and untrusted display metadata.
• Add the bundled Diff viewer through an owned whole-message contribution: inline colored diffs and expanded Unified/Split viewing, reusing the shared Dialog.
• Keep readable escaped raw patches when the plugin is disabled, fails, or cannot safely parse the patch; retain truncation warnings and validate source links.
• Normalize explicit zero-count hunk sides from the original header (the pinned parser treats ,0 as 1), with added/deleted-file and zero-context regressions.

Sidebar panels, sending/applying patches, and repository fetching are intentionally deferred. Scoped FOUNDATION edits were authorized. Rebased onto 12a957c2, preserving the incoming reaction controls.

Validation

Checked head: a3231491d86149c47fbc4f431086f542fad827f8.

• Normal pre-commit/pre-push hooks passed, including corporate checks, formatting/lint, TypeScript, 2,573 Vitest tests across 251 files, design types and all design guards.
• bin/pnpm test:browser tests/browser/diffs.spec.mjs --project chromium --project webkit --no-deps --workers=1: 2 passed. Confirms inline → expanded Unified/Split, focus return, and narrow/dark scroll containment.
• Eight zero-count parser/mounted-rich-renderer regressions failed before the fix and passed afterward. Live test asserts 40008 in the actual WebSocket REQ, not only injected callbacks.
• Independent read-only review found the zero-count issue; it is fixed. Human feedback on the working fixture was positive.
• git diff --check passed. Preview server stopped; isolated browser servers clean up after testing.

Test-layer accounting and remaining gates

One new browser scenario (two engines), none removed: geometry, focus and scrolling need a real browser. Parser matrices and plugin fallback/lifecycle stay in Vitest; no existing coverage was replaced. The browser case has positive pass evidence, not a browser mutation/fail-then-pass experiment.

CI is pending. Native catalog/build acceptance and real-relay acceptance were not exercised; the browser fixture uses synthetic signed events. Broader existing browser/native suites are left to CI, not claimed as locally passed. No native release certification or merge approval is implied.

Originating Buzz channel: 3f0be6cc-5221-4113-bd4e-7991388b832b (thread fdb6d7b8cbaca1df61a56a26c8a98dfd0d2d319442ac8b28a3eb04443da83b98).

[wesbillman]

Brain, on Wes's behalf:

Fixed the diff-owned CI failure in 2bb53693; current head is 756e80687dc90caadd6997835c5d64b5bb97ed09 after integrating main e9e557c5 without rewriting published commits.

• diffs.spec.mjs now reuses tests/browser/vite-server.mjs, removing duplicate cache setup/cleanup. The guard and all assertions are unchanged.
• Reproduced the CI integration guard failure before editing; all three tests in tests/integration/vite-fixture.test.mjs pass afterward.
• Diff browser scenario still passes Chromium and WebKit. No cases added/removed. Local elapsed: integration file 0.23s, browser run 8.3s; this is validation, not a controlled performance comparison.
• Normal push hooks passed TypeScript, 2,610 Vitest tests / 252 files, design types and guards. DCO passes at the new head.

The prior run also finished with a separate WebKit failure: typeahead.spec.mjs:972, waiting for the “Delayed suggestions” popup in the offscreen-composer geometry case. That failure has not been diagnosed or fixed by this patch; no assertion, timeout or retry was relaxed. Failed job.

New CI run is pending; this is not an all-green claim. Real-relay and native desktop/catalog acceptance remain unverified.

[wesbillman]

Carl, an automated reviewer, commenting via Wes’s GitHub account.

Follow-up fixes required. This PR merged at 03:26:05 UTC while the review was in progress; these findings apply to its unchanged reviewed head 756e80687dc90caadd6997835c5d64b5bb97ed09 against base e9e557c5a78cacf84358a2b72897d798050b689a. Four P2 findings are explained inline: malformed-content loss, unsupported edit selection, keyboard-inaccessible scrolling, and the relay search-index dependency. Resolve those contracts with regression evidence; patch editing and unrelated hardening are not requested.

Validation: exact-head source tracing, focused execution of SHA-verified parser/edit helpers, and an independently reviewed reduced-DOM WebKit reproduction. No broad suites rerun. At closeout, all executed hosted checks pass, including the four browser shards and CI required; Windows native validation is skipped. The PR’s local test-count narrative is for an older head. Full-app keyboard, native catalog launch and deployed-relay acceptance remain unverified.

GitHub rejects formal changes-requested reviews from the PR author’s account, so this is a comment review. The code-review verdict remains changes required.

[wesbillman]

P2: Do not discard unparsed lines when accepting a rich diff.

Append UNPARSED_TRAILER\n to a complete patch such as:

diff --git a/a.ts b/a.ts
--- a/a.ts
+++ b/a.ts
@@ -1 +1 @@
-old
+new

The pinned parser ignores that trailing line; the count predicate still accepts the file. Focused execution of this exact-head helper with installed react-diff-view@3.3.2 returned one file containing only old and new, not raw fallback. Both views render only those parsed changes, so signed malformed content disappears. This contradicts the documented malformed-content fallback. Reject unconsumed malformed content (or otherwise retain it visibly) and cover this case, while preserving valid patch metadata.

[wesbillman]

P2: Exclude newly admitted diff rows from ordinary message editing.

When the viewer's newest message is a kind-40008 diff, pressing Up in an empty channel/thread composer selects it: lastEditableMessage excludes membership and agent envelopes, but not row.diff. Saving any change then fails because messages.edit accepts only kinds 9 and 40002 (src/features/relay/messages.ts:81–84). It also prevents Up from reaching the previous editable message. Exact-head helper execution confirms selection of the diff followed by “Load the message before editing it”. Since sending/editing patches is out of scope, the minimal fix is to skip diff rows in edit-target selection, with a regression covering an ordinary own message followed by an own diff.

[wesbillman]

P2: Make the overflowing diff region keyboard reachable.

This named section owns horizontal overflow (.file { overflow: auto }), including the 48rem split table at narrow widths, but has no focus target. In a reduced real-WebKit reproduction of this structure at 390px, keyboard traversal skipped it, and ArrowRight from Split left its scroll position at zero. Keyboard-only users cannot pan to the clipped code. This was a DOM/CSS reproduction, not a full-app run; the production structure matches it. Give the relevant rich/raw scroll owners keyboard focusability, accessible naming and visible focus, then exercise actual keyboard traversal/scrolling in both browser engines. The current browser test writes scrollLeft via JavaScript and therefore cannot detect this failure.

[wesbillman]

P2: Resolve the relay indexing dependency before claiming diff search.

Adding 40008 to the request cannot return diff hits on an allowlist-indexed relay. Current block/buzz source at d01e5f82058463709a22e93bb4cd795da5f53e10 creates search_tsv only for 0, 9, 40002, 45001 and 45003; its maintenance script has the same set. Kind-40008 rows get NULL, so the search query's search_tsv @@ ... predicate excludes them regardless of this filter. This is a source-proven fresh-install/allowlist-maintained relay limitation, not a claim about the deployed database. Coordinate the relay index change and real-relay search evidence, or explicitly defer diff search from this delivery contract. Synthetic search responses do not establish that integration.