[TSVRN9]

Closes #80, Closes #81, Closes #82

Three related findings from PR #78's review, all on halls/[slug] and rank-surfaces.spec.ts, fixed on one branch since they share a verification loop.

#81 -- pointer-events wrapper untested (elementFromPoint gap)

The { trial: true } actionability checks already in rank-surfaces.spec.ts scroll their target into view first, so they can't tell "reachable after scrolling" from "blocked by an overlay right where it already was" -- PR #78's review found that gap: removing the compare-prompt stack wrapper's pointer-events-none class breaks click-through in reality, but every existing spec in the file stays green regardless.

• Added a desktop elementFromPoint probe: after logging a dish far down a long-menu fixture and scrolling to the document bottom, asserts the element at the last dish row's Log button center is the BUTTON, with a precondition that the probe point actually falls inside the stack wrapper's band (so a future layout change can't silently turn this into a no-op).
• Red-first: removed pointer-events-none from the wrapper -> this new spec failed (topmost element was the wrapper DIV, not the button) -> restored -> green. Confirmed under the mutation that the other 10 specs in the file stay green (matching the review's finding), and re-confirmed the mutation still goes red after web: compare prompt occludes the last dish row's Log button at narrow viewports #82's fix landed (see below), so web: compare prompt occludes the last dish row's Log button at narrow viewports #82 didn't silently neutralize this probe.
• Replaced onboarding.spec.ts:51's retired page.route("**/api/**") residency-guard form with the page.on("request") cross-origin+/api/ pattern from home-dashboard.spec.ts. Mutation-verified: temporarily injected a no-cors cross-origin fetch via page.addInitScript, confirmed the new guard catches it, reverted.

#82 -- narrow-viewport occlusion

At 375x667, the bottom-anchored prompt (min(92vw, 28rem) wide) covers the last dish row's Log button until dismissed.

• Minimal fix: an aria-hidden, sm:hidden spacer just before the fixed stack, sized to the stack's own rendered height (bind:clientHeight, reactive, no manual effect) plus the wrapper's bottom-5 offset. Only applies below Tailwind's sm (640px) breakpoint, so desktop geometry -- and web e2e: pointer-events wrapper on the compare prompt is untested (elementFromPoint gap) #81's desktop probe -- is unaffected. No new component.
• Red-first: added a 375x667 elementFromPoint spec (same technique as web e2e: pointer-events wrapper on the compare prompt is untested (elementFromPoint gap) #81's, worst case = scrolled to the document bottom with the prompt up) -- ran it against the code without the spacer and got red (topmost element was the compare-prompt's own choice button, reproducing the review's measurement) -- added the spacer -- green.
• Updated the file's ponytail: comment: removed the now-fixed narrow-viewport ceiling, added the residual one (a user already scrolled to the exact bottom when the prompt appears doesn't get the button moved to them, due to scroll anchoring -- the sanctioned "bottom padding equal to stack height" tradeoff, not a regression versus pre-fix where nothing helped).

#80 -- rank sync drift

chooseCompare() on halls/[slug] updated both Elo tracks but never called syncDiningHallRanks, per #67's original zero-server-calls constraint for that surface. #67 is now closed, and per the orchestrator's decision this is superseded: chooseCompare() now fire-and-forget syncs favorite dining halls exactly like /rank's choose() does, gated on a session existing -- sanctioned by CLAUDE.md's data residency table (favorite dining halls are the one ranking-derived, hall-level thing the server may see, signed-in only).

• Red-first: two new e2e specs using the mocked-session technique from friends-notifications.spec.ts.
  • Signed-in: mocks rest/v1/favorite_dining_halls, logs two same-hall dishes, chooses a winner, asserts both the DELETE and the POST (insert) fire -- both fixture dishes are Hampshire, so a single comparison meets MIN_RATED_DISHES_PER_HALL, proving a rank actually got pushed, not just a no-op call.
  • Signed-out: this spec is the residency guard, page.on("request") style (not route-glob counting) -- asserts zero Supabase/cross-origin requests during the same signed-out flow.
  • Ran both against the pre-fix chooseCompare(): signed-in spec failed (no sync call at all) as expected; the signed-out guard stayed green identically before and after, so it was separately mutation-verified by temporarily dropping the session && guard in chooseCompare() -- confirmed it turns red (a real cross-origin request lands) before reverting.

Verification (all local, per CLAUDE.md's "CI is local-only")

• CI=1 PORT=4317 npx playwright test (full suite) -- 70 passed
• pnpm --filter web check -- 0 errors, 0 warnings
• pnpm --filter web build -- succeeded
• shared/ untouched, so pnpm --filter @udine/shared test wasn't required and wasn't run.

One pre-existing flake observed and disclosed, not introduced by this branch: under a full-suite parallel (4-worker) run with an unrelated mutation applied, rank-surfaces.spec.ts's "home's Your Top Dishes module..." spec failed once on an element-not-found timeout. It passed 3/3 in isolation and passed in every other full-suite run in this branch's history (including the final one above). Retries are 0 by this repo's own config (playwright.config.ts), so nothing laundered a real failure -- this reads as parallel-worker contention, not a regression from this diff, but flagging it since a re-run under load could surface it again.

Per-issue evidence with the same red/green detail posted as comments on #80, #81, #82.

Not merging -- for review.