fix(daemon): surface proven no-effect gestures to the agent

[thymikee]

Fixes #1600.

What

The #1542 post-gesture stabilization loop already proves when a scroll/swipe/pan moved nothing: its accept-stale verdict fires only when the quiet post-gesture capture still equals the pre-gesture baseline after the 3.5s distrust cap — with scope-drift and root/keyboard-chrome subtleties already handled by classifyBaselineSurfaceEvidence. But the verdict only reached the diagnostics stream; the agent-facing response reported plain success.

This PR returns the verdict from capturePostGestureStabilizedResult alongside the capture value, and the snapshot handler appends a warning through the existing annotations.warnings channel (already rendered by every client path — zero new plumbing):

scroll up produced no visible change: the tree still matches its pre-gesture state. Either the container is already at its edge, or it ignores synthesized scrolls — a raw drag moves such lists: swipe x1 y1 x2 y2 (start inside the list).

The wording deliberately admits the at-edge ambiguity: iOS has no authoritative can-scroll signal (Android does), so "at edge" vs "container ignores synthesized scrolls" cannot be distinguished platform-side — but either way the agent's next move is informed instead of blind.

Why

AppControlBench element-18 (gpt_high): reaching "Manage sessions" in Element's settings table cost 433s / 80 tool calls, ~40 of which were scroll/fling/pan attempts that all reported success over a byte-identical tree — only a raw swipe moved the list. The daemon's stabilization loop (post-#1542) detects exactly this condition; it just never told the agent.

Red evidence

Both new tests fail on base 3696f338b / current main pre-fix:

× capturePostGestureStabilizedResult keeps polling past the normal deadline when the AX tree is stuck at the pre-gesture baseline (iOS)
  → result.gestureNoEffect is undefined (verdict never surfaced)
× formatGestureNoEffectWarning names the gesture and the raw-drag escape hatch
  → function did not exist

Live verification (seeded bench Bluesky feed, worktree build)

action	ground truth	response
scroll up 1 at feed top	no movement (at edge)	warning ✓
scroll down --pixels 600 mid-feed	first item whiskers→buddy	no warning ✓
raw fling at feed bottom	first item unchanged (verified)	warning ✓

Notes

• positionals stored on the stabilization marker feed the wording only (numeric args filtered out); nothing is re-dispatched.
• Android is untouched: its captures are fresh by construction (Android snapshot stays stale after the second Navigation 3 navigation (in-ComposeView content swap) #1254), so no baseline distrust and no stale-accept verdict exists there.
• The interaction-outcome-aware path (tap + trailing gesture) surfaces the same warning through the same helper.

[github-actions]

Size Report

Metric	Base	Current	Diff
JS raw	1.96 MB	1.96 MB	+1.0 kB
JS gzip	627.1 kB	627.4 kB	+328 B
npm tarball	748.7 kB	749.1 kB	+331 B
npm unpacked	2.62 MB	2.62 MB	+1.0 kB

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	27.0 ms	28.1 ms	+1.1 ms
CLI --help	63.7 ms	66.9 ms	+3.2 ms

Top changed chunks:

Chunk	Raw diff	Gzip diff
dist/src/internal/daemon.js	+1.0 kB	+328 B

[thymikee]

P1: Do not present accept-stale as proof that the gesture had no visible effect. classifyBaselineSurfaceEvidence is subset-tolerant: unchanged shared chrome can produce unchanged while newly visible list cells are absent from the baseline and ignored. #1573 now has live evidence of the same terminal stale_accept after content demonstrably moved. This PR would turn that known false negative into the agent-facing claim ‘tree still matches its pre-gesture state’ and recommend a different gesture. Keep the diagnostic heuristic internal, or require stronger full-surface evidence (including unmatched discriminating entries) before emitting a no-effect warning, with a fixed-chrome + replaced-list regression test.

[thymikee]

Addressed in 7af45ba — the P1 was correct, and the fix follows the "stronger full-surface evidence" option:

• The claim is no longer keyed off accept-stale alone. gestureNoEffect now additionally requires haveIdenticalDiscriminatingSurfaces(baseline, quiet): every discriminating entry must match exactly, in both directions, rects within tolerance. Any appeared or vanished real element — the fixed-chrome + replaced-list shape, and also plain scope drift between baseline and capture — vetoes the warning. The subset-tolerant classifier stays untouched for the distrust loop, where a false 'unchanged' only buys extra polling.
• Added the requested regression test: fixed tab-bar chrome + fully replaced list cells → the loop still accepts stale (staleAccepts: 1, loop semantics unchanged) but gestureNoEffect is absent. Red on the previous PR commit (the claim was wrongly present), green with the gate.
• Corrected the overclaiming docs ("PROVABLY") to state the corroboration requirement explicitly.

Trade-off accepted: the warning becomes rarer (scope-drift no-ops now stay silent), which is the right direction for a message that steers the agent's next gesture. The live-verified cases from the PR body (top-edge scroll, bottom-edge fling on the seeded feed) still warn — identical presentations on both sides.

🤖 Addressed by Claude Code

[thymikee]

Re-review at 7af45ba3c: the prior false no-effect finding is resolved. The agent-facing claim now requires equal discriminating-entry sets in both directions plus rect agreement, and the fixed-chrome/replaced-list regression correctly stays accept-stale internally while withholding gestureNoEffect. No further code finding.

Not ready on exact-head evidence: Fallow deterministically rejects capturePostGestureStabilizedResult at cyclomatic 14/cognitive 16/68 LOC, so extract the acceptance/result construction; iOS Smoke also timed out in the native-alert scenario and must be rerun/isolated before classifying it as infrastructure.

[thymikee]

Re-review at 75492bf55: the complexity extraction is behavior-preserving; the full-surface corroboration conjunction and regression remain intact, and exact-head Fallow is now green. Code review is clean. Marking ready for human review while the remaining exact-head smoke/coverage/integration lanes complete.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-08-04 19:23 UTC