fix(android): demote shared android:id/* framework ids below a discriminating label (#1269)

[thymikee]

Summary

Fixes #1269.

On Android, replay of native list UIs produced systematic false divergences (FDR 1.0 measured on Settings flows) because the record-time annotator chose the non-unique framework resource-id android:id/title as primary identity for list-row text — it matches every titled row on a screen (11 matches on the Settings root in the reported repro), and the label-less generic container ancestry (linearlayout/recyclerview/framelayout) gives disambiguation nothing else to work with but sibling/viewportOrder position.

Tracing the recorded expression end to end confirmed the exact mechanism:

• The human-facing selector chain led with id="android:id/title", which at replay matched all 11 rows. resolveSelectorChain's own in-clause disambiguation (depth/area heuristic) resolved a winner from that same ambiguous clause and never even reached the later role=... label=... clause that would have resolved uniquely.
• The target-v1 binding metadata recorded id: "android:id/title" as primary identity, so matchesLocalIdentity treated any node sharing that id as sharing identity (ignoring label), leaving only sibling/viewportOrder to disambiguate.
• When a row shifted (a conditional row inserted, or the shared-id set reordered), both mechanisms picked the wrong element, and the identity verifier correctly — but falsely — refused with identity-mismatch.

Fix

This is a record-time identity-selection fix, not a verifier change — the verifier's refusal was doing its job (a safe failure, not a silent mis-bind), so it stays untouched.

A shared android:id/* id is now demoted below a recorded label whenever a label is available to discriminate instead:

• src/replay/target-identity-node.ts — readNodeLocalIdentity (the ONE local-identity reader shared by the record-time writer, replay-time verification, and the dispatch post-resolution guard) drops a non-discriminating android:id/* id from the recorded LocalIdentity when a label exists, so target-v1 verification narrows its identity/disambiguation domain by label instead of by the shared id. Adds the isNonDiscriminatingAndroidId predicate.
• src/selectors/build.ts — buildSelectorChainForNode reorders the recorded, human-facing selector chain (id=... || role=... label=... || label=...) so the discriminating role=... label=.../label=... clauses lead; the shared id survives only as a last-resort fallback entry at the end of the chain, so replay resolves via the (unique) label instead of the ambiguous id.

Both call sites share the one predicate so the recorded expression and its target-v1 metadata stay consistent. A unique, app-namespaced resource id (e.g. com.android.settings:id/network_and_internet_row) is unaffected and still leads the chain/identity. A shared framework id with no label to fall back on is also left untouched (nothing else to discriminate with).

Tests

New unit tests pin:

• a shared framework id (matchCount 11, matching the reported repro) is not selected as primary identity when a label exists (src/replay/__tests__/target-identity-node.test.ts, src/daemon/__tests__/session-target-evidence.test.ts, src/selectors/build.test.ts)
• a unique (non-framework) resource id is still preferred (same files)
• a shared framework id with no label is kept, since nothing else can discriminate
• an end-to-end regression reproducing the issue's exact scenario: record a row among 11 sharing android:id/title, insert a conditional row ahead of it (the position shift from the issue), and confirm the recorded selector-chain expression still resolves the correct row via its label clause with the id demoted (src/daemon/handlers/__tests__/session-replay-target-classification.test.ts)

Test plan

• pnpm test (full unit suite): 4064/4066 passing. The 2 failures (session-replay-vars.test.ts Maestro runScript http tests) are a known timing flake under full-suite CPU contention — confirmed passing in isolation (npx vitest run --project unit-core src/daemon/handlers/__tests__/session-replay-vars.test.ts → 75/75 passing).
• pnpm check:layering — the new selectors -> replay import edge (an unranked zone) is clean, no cycle.
• npx tsc --noEmit — clean.
• oxlint on changed files — clean.
• Live-device FDR re-measurement on Android (wave-3 E1-style repro) — pending, not attempted in this environment per the issue's verification guidance (no live device/emulator available here).

[github-actions]

Size Report

Metric	Base	Current	Diff
JS raw	1.7 MB	1.7 MB	-26.3 kB
JS gzip	557.5 kB	550.7 kB	-6.8 kB
npm tarball	673.0 kB	665.1 kB	-7.9 kB
npm unpacked	2.4 MB	2.3 MB	-31.0 kB

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	27.3 ms	28.8 ms	+1.5 ms
CLI --help	59.6 ms	59.0 ms	-0.6 ms

Top changed chunks:

Chunk	Raw diff	Gzip diff
dist/src/sdk-selectors.js	+22.0 kB	+7.2 kB
dist/src/session.js	-25.6 kB	-6.6 kB
dist/src/tv-remote.js	+14.1 kB	+5.1 kB
dist/src/internal/daemon.js	-3.7 kB	-1.1 kB
dist/src/registry.js	-1.6 kB	-541 B