test(artifacts): fix CI-breaking fixture gap + post-merge review findings

[bermudas]

Summary

Post-merge review follow-up on the artifacts module (TC-030..043, merged in #123). Fixes a CI-breaking gap plus several review findings — no new test cases, no rescoping.

Critical — CI could never actually run this suite:

• tests/artifacts.spec.ts sourced its 24 fixture files from the gitignored, local-only Elitea-testing-WebQAPreExecuted/Elitea_test_data/artifacts/ directory — every fixturePath() call resolved to a path that doesn't exist on a fresh checkout, so @artifacts could never run in CI at all.
• .gitignore also carried a case typo (webQAPreExecuted vs. the real WebQAPreExecuted) — harmless on macOS (case-insensitive git config) but the pattern was wrong regardless; fixed.
• Fix: established tests/fixtures/artifacts/ as this project's first committed-binary-fixture convention, copied the 24 needed files there (SHA-256-verified byte-identical to the originals), and updated fixturePath()/FIXTURES_DIR to point at it. actions/checkout alone now makes them available in CI — no separate provisioning step needed.
• test-large-image.png (27MB, TC-033 only) was not committed — replaced with an in-memory buildOversizedImagePayload() helper that pads the already-committed test-image-small.png to 4MB and hands Playwright a FilePayload buffer directly (the app's size-limit check is confirmed client-side/byte-length-based, so this is functionally identical to the original fixture for this test's purpose).

Important fixes:

• Strengthened 5 assistant-reply-content assertions (TC-030/035/039/040/041) that only checked toBeVisible() or a trivial /\S/ regex, to match what their own AFS documents as "the strongest available proof the model genuinely processed the image" — using the literal text embedded in each fixture (generate_test_images.py) as the assertion target.
• TC-035's chat-modal GH#114 check compared an <img src="blob:..."> attribute before/after a wait, which can never detect GIF frame animation (a blob URL doesn't mutate on frame change). Replaced with the same .screenshot() pixel-diff technique the sibling Artifacts-bucket-panel check two steps later already uses correctly.
• TC-034: implemented the AFS's claimed size/bounding-box comparison (modal image genuinely renders larger than the inline thumbnail) instead of a bare toBeVisible().
• ArtifactsPage.messageThumbnail() is now row-scoped via an optional messageText param, matching its siblings assistantReply()/attachmentFileCard() — previously page-wide, only safe because startNewConversation() happens to guarantee a cleared thread.
• TC-036: added the AFS's download.failure() → null assertion, and added the assistant-reply wait that two sibling doc comments already (incorrectly) claimed TC-036 performed.
• TC-043 previously leaked 10 files into the shared bucket every run with no cleanup. Added real teardown (same deleteArtifactAndVerify() pattern as TC-042) and amended the AFS's Cleanup section to document why the original "no cleanup" call doesn't transfer to a real storage footprint.

Nits:

• Added the previously-missing attachCounterText() check to TC-032 (matches its own AFS).
• Standardized the // Known defect: GH#NN source-comment convention for TC-034/GH#119 and TC-035/GH#114 (previously only in the assert message/step title).
• Deferred: TC-030's attachCounterText() (would require touching the shared attachFileAndSend() helper used by 4 other tests) and TC-033's bucket-totals corroborating check (low value given the network-level proof already covers it).

Test plan

• npx playwright test tests/artifacts.spec.ts run locally at least twice — all fixes verified functioning; TC-034/TC-035 fail for the known, already-filed, non-blocking product defects GH#119/GH#114 (expect.soft(), unchanged from the merged baseline)
• CI run triggered via gh workflow run playwright.yml --ref tests/artifacts-module-postmerge-fixes — confirms the fixture-provisioning fix specifically (fixtures load from the committed tests/fixtures/artifacts/ path, no missing-file errors)