# Round-2 iOS recapture replay

Evidence-only scripts; no product edits. Before 23075af981e40f7c1071871782d489896c3919f6; after 7c76735fd24aa08881422e73bf2d5ac0484ef7d3. Use detached checkouts /home/ubuntu/wt/pr-29075-recap-ios-{before,after}, vp install, and mobile:install ios --prepare for each. Both have native fingerprint 887f84d7e8ee2e1063658bf884c9d17f60d6cbb3. All servers start with T3CODE_DEV_SEED=0 and --host 0.0.0.0. No live database is read or copied.

1. Start vp run dev in each checkout. This run used core before 14844, after 15809; read dev-runner actual ports and adapt scripts. Start external servers using vp run dev:server --host 0.0.0.0 --port 16141/16142 --home-dir <scratch>/server-before/server-after --auto-bootstrap-project-from-cwd=false. Issue private auth session tokens using each revision's CLI. Never upload tokens or pairing/startup logs.
2. Run seed.mjs side for core. For external override IOS_FIXTURE_PORT, IOS_FIXTURE_HOME and IOS_TOKEN_FILE. Run ensure-scratch.mjs then seed-scratch.mjs side. The real projects and stopped sentinel threads come from production CLI/RPC; no provider turns run.
3. Connect the device pool from the before checkout with unique sessions pr29075-recap-ios-phone/tablet. Lease exactly iPhone 17 Pro and iPad Pro 13-inch (M5); install through mobile:install ios --session. Rotate iPad landscape-left. Start Metro with vp run dev:client --port 8091 from apps/mobile in the revision being captured (sets APP_VARIANT=development and scheme t3code-dev); wait for /status before opening the native development-client URL. Use the VM Tailscale IPv4 if DNS fails. Pair via the repository pair-client.sh using fresh credentials. Accept visible native Open confirmations. After pairing, relaunch the development-client URL (without a pairing URL) before a JS reload; otherwise the consumed pairing link can reopen an invalid Add environment sheet.
4. F03/F04: isolate to the core environment with remove-all-connections.js and pairing. Select Bravo in the actual picker, Back on phone; clear workspace selection through persisted preferences and reload on iPad. reset-core-remembered-bravo.js restores that same real Bravo reference when recovering a capture. Freeze relativeTime formatting with freeze-ages.js, without changing Date.now, auth, selection or RPC. Reapply after JS reload. Run capture-core.py session prefix phone|tablet. It records All -> project filter Alpha -> Alpha sentinel -> global compose -> hero -> Back/filter retention (iPad leaves detail visible).
5. F10: isolate to external. Remember its Bravo; apply-legacy-scratch.js persists the real scratch workspace key; await preference-result.js Success, reload, freeze ages. capture-scratch.py side records legacy No project -> global compose -> hero -> Back -> explicit canonical No project -> compose -> canonical hero -> retained filter. It uses the actual scratch project and sentinel, not a fixture-only visual.
6. Multi-environment rows: after core/scratch captures, run seed-multi.mjs side e1 and e2, then order-multi.mjs side. If a server exits after project creation, restart it against the same isolated home and resume seed-multi.mjs with IOS_SKIP_PROJECT_ADD=1; do not recreate or repair its database. Pair both external E1 and core E2. rename-fixture-environments.js uses the production connection update command and real IDs. seed-remembered-alpha-e1.js persists remembered real root Alpha@E1 with empty selection and filters, repository grouping/light; seed-long-path-dark.js uses the same remembered root with repository_path grouping/dark. Await preference-result.js Success, reload through a Metro launch URL, reapply freeze-ages.js.
7. capture-boundary.py side F14 records All -> environment E2 -> repository Alpha scope -> compose -> hero -> environment selector. capture-group.py side F18 <long project title> records All -> long apps/web path scope -> compose -> hero -> environment selector. Both save real scopes and React route draft composite (environmentId, projectId) through read-only CDP traversal in draft-target.js.
8. F08: only one hardware injection attempt: agent-device key n --modifiers cmd --session pr29075-recap-ios-tablet. CLI 0.21.32 returned INVALID_ARGS Unknown flag: --modifiers before injection. No callback simulation is substituted in this recapture.
9. capture.py takes full-screen screenshots 500 ms apart. Final hero stills use --normalize-status-bar on both sides. Some normalized captures fail full-frame equality across ten pairs because the charging indicator animates; others eventually stabilize. See capture-stability.json for the actual result per accepted frame. If the final pair is pixel-identical below the native status area it saves that full-screen still and records the exact excluded rows: iPhone y=0..61, iPad y=0..31. compare.py computes both full-frame Diff0 x32 and DiffT >16/255, plus status-masked equivalents. No home-indicator/bottom rows are masked; retake unrelated app-region differences. Recordings retain real clocks and show normalization toggles during the hero capture window; they are not used for pixel comparison. Thread-age formatting is fixed to 2026-11-05T00:00:00Z, although initial cached unfiltered labels can retain elapsed ages until filtering rerenders them; compared iPad hero ages are fixed at 1d on both sides.
10. contact-sheets.py decodes original MP4s for inspection. Open accepted full-screen captures and all diffs before ghdrop. upload-artifacts.py preserves ghdrop -m stdout verbatim. Public artifacts exclude credentials, pairing screens, private logs and rejected attempts. Stop only owned port processes after confirming cwd, close leases, remove only own two worktrees. Keep this scratch directory for lead handoff.

Accepted prefixes are recorded in accepted-runs.json. Rejected attempts are preserved locally, excluded from the evidence uploads. F10 after recording has an initial native reload/loading interval before its first settled filtered still; the matching compose and canonical-filter steps follow. Video command offsets are approximate and are not pixel-comparison inputs.

For the additional phone retake, session pr29075-recap-ios-retake leased device 29900e15. Set IOS_DEVICE_NAME_MATCH=29900e15 for CDP; set REACT_NATIVE_PACKAGER_HOSTNAME=fs-21.tail131df4.ts.net when starting Metro so the CDP Origin is accepted. capture-group.py accepts optional session and output-prefix arguments after its long-title argument.

F18 was repeated on the same new phone lease on both revisions. Despite matching launch context, two one-level red/blue pixels at (36,62) and (39,62) remain outside the hero title and below the permitted status mask. Both capture attempts are retained locally; the final r2 pair is uploaded with comparability explicitly unverified. Do not expand the mask or claim zero-threshold cleanliness for F18.
