Map: a verified go/no-go for 2.4.0 — every feature area on the Pixel 6 before develop meets main

[simonoppowa]

Destination

Reached 2026-09-26. 2.4.0+66 shipped: release PR #1270, v2.4.0, Play internal via the API and TestFlight. Store promotion and review submission remain by hand, past this map.

A verified go/no-go for 2.4.0+66. Every feature area has a verdict from the Pixel 6; each failure is triaged against the bar; changelogs/66.txt is written and the two pre-PR checks from docs/RELEASING.md are done — so the next step is opening the develop → main PR with nothing left to remember. The PR itself, the release-ship approval and the store consoles are past the destination. Unlike most maps here this one carries execution: building, installing and running the test are task tickets on the route, because the verdict cannot exist without them.

Notes

Domain. Flutter app, conventions in AGENTS.md; docs/RELEASING.md is the checklist this map feeds — a release is a merge of develop into main, and the pipeline does the rest. Sessions reach for /grilling and /domain-modeling on the triage and changelog tickets. Task tickets that need the phone in hand are HITL: the maintainer taps, the session hands over a script and records what came back.

The candidate. develop at 86cd09d9 (2026-09-19), frozen. Both main and develop read 2.3.0+65; v2.3.0 shipped 2026-09-10 as build 65. The real delta is the diff of lib/ against the v2.3.0 tag (89 files) — git log main..develop is polluted by squash-merge and lists the 2.3.0 release's own commits. What changed, by area:

Area	Since v2.3.0
add-meal search	food names requested in the app's language, not the device's (#1215); Spanish and Hungarian names from Open Food Facts and the backend (#1213, #1229); canonical FDC sibling + full backend descriptions (#1170); has_portion read at the search cut (#1209); scan / create-custom actions on the no-results state (#995)
AI meal assistance	the model's portion word matched on the English label, bare ties broken toward the middle rung, misses marked (#1208); elapsed time, hint and cancel while an own server answers (#1149)
Open Food Facts data	micronutrients converted into app units on new writes (#775); niacin read from vitamin-pp_100g (#1168); a startup repair rewrites OFF rows logged before #775 (#1193, OffMicronutrientRepair, keyed on MealDBO.dataVersion)
meal detail	quick-serving chips 0.5×/1×/2×/100 g (#1101); opt-in Settings toggle for grams-first default (#1195); hydrated record kept across the unit menu (#1221)
profile	7-day moving-average overlay on the weight-trend chart (#1197)
recipes	totalWeightOverridden threaded through JSON and CSV import (#1169, #1196)
settings	export/import failures logged and explained to the user (#1150); grams-first toggle; Spanish + Hungarian in the language picker
l10n	Spanish and Hungarian shipped; Weblate pipeline live (#1171), 15 locale files updated
deps	hive_ce 2.19.3→2.20.0, mobile_scanner 7.4.0→7.4.1, sentry_flutter 9.26→9.30, cached_network_image 3→4, archive 4.0.9→4.2.0, flutter_local_notifications 22.3.0→22.3.1
icons	launcher artwork inset inside the adaptive mask (#1161)

Untouched since v2.3.0, smoke depth only: diary, fasting, add-activity / activity detail, Health Connect, scanner UI, onboarding (on the candidate, after pm clear of .develop), theme, notifications.

Known issues the candidate carries — recorded here so a tester does not re-find them, and none is a blocker under the bar: the weight-trend chart windows by 24-hour spans, so a DST change inside the window drops a day (#1207, fix in #1222); several activity picker rows share an English name (#1200, fix in #1223); a cleared per-app language override is re-pushed on the next cold start (#810, fix in #1230). All three PRs wait for the next release.

The phone. Pixel 6 (oriole, 18101FDF60010M) on Android 17, running the app in German. Its history matters: at charting it held full 2.2.0 with the maintainer's real data and .develop 2.3.0; both were wiped on 2026-09-19, so the 2.2.0 real-data upgrade is gone for good. It now carries one install — full, first installed 2026-09-19 as v2.3.0, upgraded in place to the candidate on 2026-09-25 — over a box holding two Open Food Facts intakes and the product cache, and nothing else. Since 2026-09-25 the session drives the phone itself via adb, re-checking mCurrentFocus after every event and stopping if it leaves the package; the system file picker during export is the one expected hand-off. The maintainer is still needed for the AI provider key and a barcode.

Builds. Debug-signed (0bdae1d6…5ff5), which is the key every install on this phone carries, so adb install -r is a true upgrade. Recipe: scratch worktree with .env copied in, then flutter pub get, flutter gen-l10n, dart run build_runner build --delete-conflicting-outputs, flutter build apk --flavor full --debug — about four minutes. The scratchpad is tmpfs: APKs and worktrees do not survive a reboot and have been rebuilt twice; durable evidence goes to ~/Documents/ont-release-2.4.0-evidence/.

Evidence. The exports (ont-develop-230.zip, ont-full-220.zip, pulled to the session) are the numeric baseline for the #1193 repair — the screens round pre-#775 values to 0.0 — and the repair recomputes a recipe's aggregate while leaving recipe-sourced intakes alone. Scripts live on research/release-2.4.0-pixel-test under docs/release-tests/2.4.0/.

Reporting. A HITL run reports per area: pass / fail (what you did, what you saw, what you expected) / skipped (why). The ticket's resolution comment holds that table; the triage ticket reads it.

Decisions so far

• What is the release candidate, and which version does it become? — develop at 86cd09d9, frozen; it becomes 2.4.0+66. The five open PRs wait for the next release; the three bugs they fix are known issues of the candidate, not blockers. The bump commit is the only thing allowed on top.
• How does the candidate reach the Pixel, and who drives the test? — two local full --debug builds, v2.3.0 first, the candidate over it; the maintainer drives. Same debug certificate makes adb install -r a real upgrade, so hive_ce, the DBO changes and the fix(data): bring Open Food Facts micronutrients logged before #775 into app units #1193 repair run on a v2.3.0 box. The session installs and reads logs, sends no input.
• How wide is the test, what blocks the release, and is AI assistance in the run? — every area at smoke depth, the delta in depth; crash, data loss, wrong nutrition numbers or a v2.3.0 regression block; AI against one provider. Everything else ships with an issue filed.
• Build the v2.3.0 and candidate debug APKs, and install v2.3.0 on the Pixel — three debug APKs built (candidate as develop and full, v2.3.0 as a spare), nothing installed. The target is a Pixel 6 on Android 17 already carrying .develop 2.3.0-develop (65) and full 2.2.0 (63), both on this machine's debug key; the route became upgrade .develop, then full on real data, then onboard on a cleared .develop.
• Write the baseline script: what to note on the two installs before the upgrade — written, on research/release-2.4.0-pixel-test; the exports are the numeric baseline, not the screens. Do → Expect → Note over both installs; six accidental-write paths named; .develop gets reinstalled as true v2.3.0 first because its build has fix(add-meal): convert Open Food Facts micronutrients into the app's units #775 without the stamp; the recipe aggregate is recomputed by the repair (map corrected).
• Run the baseline pass on both installs, then upgrade-install the candidate — the fix(data): bring Open Food Facts micronutrients logged before #775 into app units #1193 migration passes on the device: 0.72 → 720.0 mg on a real pre-fix(add-meal): convert Open Food Facts micronutrients into the app's units #775 row, zeros and macros untouched, stamped, idempotent, no crash. Plus 25 green unit tests. But both installs had been wiped, so the fixture is two intakes and the 2.2.0 → candidate path is unreachable — the 63 → 66 upgrade is covered by unit test only, which the go/no-go must rule on.
• Write the candidate test script: the delta in depth, every other area at smoke depth — written for the session to execute over adb, 642 lines, on research/release-2.4.0-pixel-test. 27 adversarial findings applied, including an export-filename blocker and three near-miss data hazards; fix(add-meal): convert Open Food Facts micronutrients into the app's units #775/fix(off): read niacin from vitamin-pp_100g, not niacin_100g #1168 were silently uncovered and now are; fix(add-meal): read has_portion at the search cut for the resolver's page, so a portionless exact title no longer pushes the survey record out of the twenty #1209 is recorded as not verifiable on device. Seeding is step 1, and a fourth verdict, inconclusive, exists for evidence that does not decide.
• Seed the candidate, then run the candidate script on the Pixel and report per area — run complete; nothing hit a "Blocks if". Every delta item verified on the device or in the packaged build: fix(search): request food names in the language the app is read in, not the device's #1215 (langs on the wire), fix(add-meal): resolve to the canonical FDC sibling, and show backend foods by their full description #1170, fix(add-meal): convert Open Food Facts micronutrients into the app's units #775/fix(data): bring Open Food Facts micronutrients logged before #775 into app units #1193, feat(profile): 7-day moving-average overlay on the weight-trend chart (#1119) #1197, feat(meal-detail): quick-serving chips (0.5x/1x/2x + 100 g) #1101, fix(meal-detail): keep the hydrated serving option when the unit menu opens (#1216) #1221, feat(meal-detail): opt-in Settings toggle for grams-first default (#1126) #1195, fix(recipes): thread totalWeightOverridden through JSON import save #1169/fix(recipes): thread totalWeightOverridden through CSV import and guard the column (#1194) #1196, fix(settings): log export/import failures and tell the user why #1150, feat: add scan and create-custom actions to no-results search state (#578) #995, feat(add-meal): match the model's portion word on the English label, break a bare tie toward the middle rung, and mark a miss #1208, fix(icons): give the launcher artwork room inside the icon mask #1161, the CSV export shape and the cached_network_image 4.0 bump. Three gaps left for the go/no-go: the scanner preview is inconclusive (black in a dark room), the custom meal is blocked (form not agent-driveable), and feat(bulk-add): show elapsed time, a hint and cancel while an own server answers #1149's own-server indicator needs infrastructure.
• Does any change since v2.3.0 add or alter a network destination? — no: same hosts, same RPCs, same payloads; the declarations stand with no console edit. langs= and fields= changed value only (asking Open Food Facts to return more, not sending more); every AI transport file is untouched; Sentry's gate and config are unchanged; the Android manifest has zero diff. One honest limit: whether a bumped SDK turned on a new default cannot be proved from the repo.
• Which main-only lines are real gaps in develop? — none; nothing to cherry-pick. Of ~95 files with main-only lines, 82 match the merge base outright and the 13 independent ones all check out — the Weblate hotfix pair is byte-identical, the ARB diffs are key reordering (zero main-only keys), and develop is newer everywhere else. Turned up two non-gaps that matter: develop is still at 2.3.0+65, and RELEASING.md's Notes section misdescribes the workflow sets.
• What does changelogs/66.txt say? — written as six bullets, 796 chars, in PR #1261. The file already existed (feat(add-meal): match the model's portion word on the English label, break a bare tie toward the middle rung, and mark a miss #1208's portion bullet, kept and trimmed). The micronutrient correction gets its own plainly-worded line because it rewrites numbers users already logged; plain 0.5x/hyphen because the file is pasted into the consoles by hand.
• Go or no-go: rule the Pixel's report against the bar — NO-GO as frozen, and not because of the test. Nothing on the device hit a Blocks if and both pre-PR checks are clean; but #1254 (a custom meal with base quantity 0 stores NaN and blanks Home, Diary and Trends) was reported after the freeze, and its fix #1259 is green. Verified it is not a regression — edit_meal/ is unchanged since v2.3.0 — so the bar's fourth clause misses it; ruled in anyway rather than knowingly shipping data corruption. Re-freeze scope is fix(edit-meal): stop a zero base quantity from storing NaN nutriments #1259 only.
• Re-freeze with the NaN fix, re-test the custom-meal path, and settle the scanner — re-frozen at 680eafaa (2.4.0+66); clean on device; scanner passes. Both startup repairs behaved correctly (silent no-ops), and a full export scan found zero non-finite values. The zero-base-quantity path itself was not exercised on device and is accepted on fix(edit-meal): stop a zero base quantity from storing NaN nutriments #1259's unit tests, which reproduce the reported case by name.

Not yet specified

Nothing — the map closed at its destination. The three patches still open at the end were settled on the way:

• The grams-first toggle shipping untranslated — ruled at Go or no-go: ships, as the Weblate pipeline working as designed. Its semantics identifier was added in #1270 / #1271 after the Codex review.
• The lost 2.2.0 install — accepted: the 65 → 66 upgrade is verified on device, the 63 → 66 jump by the migration's unit suite only.
• Health Connect on Android 17 — the smoke passed (screen renders, disclosure before any system UI); the permission grant stayed the maintainer's and was not exercised.

Out of scope

• Opening and merging the release PR, and approving release-ship. docs/RELEASING.md scripts it; the map ends where that checklist begins.
• Store console work — listing text, the Play data-safety record (the re-check here only decides whether it changed), the Health apps declaration, App Store submission and the Android hand-upload fallback.
• The flutter_secure_storage 11 bridge (fix(storage): move the Android keystore entries out of the options flutter_secure_storage 11 removes #1210 / chore(deps): bump flutter_secure_storage from 10.3.1 to 11.0.0 #809). It touches the AES-key path and gets its own release; the keystore migration is deliberately not on this candidate and not tested here.
• iOS. No device on hand; TestFlight is the pipeline's job and App Review reads the bundle.
• The deferred fix PRs fix(profile): window the weight-trend chart by calendar days, not 24-hour spans #1222, fix(l10n): give every activity row its own English name #1223, docs: correct the three one-liners #991 still lists #1225, fix(l10n): stop re-pushing a per-app language override the user cleared (#810, carried) #1230 — next release.