[obvious-autobuild]

Human author: Sam Gu (sgu07966@gmail.com)

What changed and why

This branch is the integration join: one branch cut from main that merges every overhaul slice of the burst5/6/8 waves, so the app, backends, data pipeline, and evaluation tooling can be verified and reviewed as one tree instead of twenty-five parallel stacks. The merges on the branch (each a merge: integrate ... commit) carry PRs #146–#180 lineage: the four spec/contract slices (#146 taste, #147 lechef, #148 recipes, #150 community), the taste store/lab/live slices (#155, #165, #167, #176), the LeChef cook/plan/tools/live slices (#152, #153, #156, #157, #171), the recipe pipeline/originals/integrate/bundle-v2 slices (#158, #162, #163, #168, #148/#151 hydration), the community client/groups/api/e2e/contract slices (#159, #166, #179), the catalog recommend/cook-flow/reach slices (#172–#175), the weekly plan (#178) and substitution canonical table (#180), plus the overhaul trunk (#170). No history was rewritten; every slice arrives as a merge of its reviewed PR head.

On top of the merges, this retry round repairs the four places where the join had drifted or failed:

• c34b82a8 — TasteEventStoreTests.swift had a committed conflict marker from the trunk port; resolved in favor of the stronger assertions (each event asserted present, failing if either is missing).

• 51f14e11 — the taste-store-check Linux mirror now pins the catalog feature index and its pure FeatureLogic inputs via a mirror-local bundle-URL shim, so the suite runs on Linux without pulling non-portable AppDatabase dependencies.

• 5a82b961 — the taste-lab reference engine was still on pre-e95e2644 semantics and failed contract parity after the spec branch's per-kind payloads and evidence gates landed: the recipe_rated merge into an unrated cook now adds the rating as its own cook_rating evidence at the merge instant (instead of replacing the no-rating update), saves use reward 0.75 and skips 0.15 (new REWARD table matching the contracts checker), the per-recipe learnedContribution reports the clamp-bounded value with pins counted in, pins render alongside inferred statements (the old pinned-opposite suppression is gone), and the hedged wording rule requires numerator == denominator. Parity tests replay the contract fixture's seeded pin statements; the clamp test's unclamped-sum oracle sums per-feature contributions. The three golden household fixtures were regenerated over the merged bundle-v2 catalog.

• edd5cc8d — evaluation/catalog-reach/reach-after.json|md regenerated for the merged state: the committed pair predated the final bundle regen (763 recipes, 461 original-rights, 0 legacy_id-resolvable keys) while the merged bundle ships 762 recipes (all original-rights), 49 legacy_id-resolvable keys, and 655/762 (86.0%) fully resolvable. reach-before is untouched.

• 26871886 — the community-check migration mirror refresh (the two drifted source pins + HASHES.txt) committed after being verified locally in the previous round.

• d5ad8965 — reconciled the parallel contract-alignment head (5a82b961): its engine, tests, and regenerated fixtures were adopted wholesale (its hedged rule — hedged only when numerator == denominator — is the faithful checker wording), the conflicts were resolved in its favor, and the one rule it had not ported was restored on top: profile staleness now decays on the feature's most recent evidence class (HALF_LIFE[post.last_class]), matching check.py's HALF_LIFE[final[fk][3]].

• 137eed07 / 65928425 — the production Swift residual engine (apps/ios/FeatureLogic/Taste/Engine/TasteResidualEngine.swift) now reproduces the committed golden household fixtures: 137eed07 landed the ingest half (the rated-cook merge adds the rating as its own cook_rating evidence at the rating's instant with the cook's exposure share instead of replacing the no-rating update; saves apply reward 0.75, skips 0.15), and 65928425 reconciled it in a rebase and ported the remaining rank/profile half — rank() reports the clamped learned sum with pins counted as learnedContribution (the same number totalScore adds), inferred-statement staleness decays on the feature's most recent evidence class, hedged low-rating wording requires numerator == denominator below 10 opportunities, and inferred statements render alongside active pins (the pinned-opposite suppression is gone). testLateRatingMergesIntoCookWindow now pins the add-on-top merge arithmetic (alpha 1.295 / beta 1.03 for a 5-star rating over a 0.25-share cook), and the taste-store-check mirror pin was refreshed. Fixtures untouched — they are the spec.

• 86bcc382 — fixed the snapshot-load reentrancy fatal behind the repeated GRDB "Database methods are not reentrant" crashes that killed LeChefPlanPersistenceTests runners: PlanDataAdapter.loadSnapshot() ran the reviewed-substitution loop inside its own dbQueue.read closure while SubstitutionService.substitutions(for:) opened a second read on the same serialized queue (bridge() → loadBridge(db: DatabaseQueue)). The service now has a substitutions(for:dietaryRestrictions:in:) overload that fetches the identity bridge on the caller's held connection (loadBridgeRows(db: Database)); the queue-based loader keeps its own transaction for standalone UI callers, and offerableSwaps takes the bridge as input so every path shares one read.

• b16e0984 — the catalog integration test pinned the legacy omelette as the quick suggestion for a tomato-only pantry, which held while the near-match pool was legacy-only. Under mergedResults' shared ranking (score desc, missing count asc, time asc — no legacy preference), the regenerated 762-recipe bundle's catalog entries legitimately compete, and a catalog recipe with a stronger health score now tops the near match (its nil legacy id surfaced as "nil" because Recipe.id is Int64?). The omelette's presence and exact fit lines stay asserted; the quick-suggestion check now pins the published shape — quick suggestion equals the top near match when nothing completes — so it guards engine/UI consistency without pinning a bundle-specific winner.

The two repairs merged with the parallel taste-engine port (65928425) as head 8dc9329f; no file overlap, merge is clean.

• Re-run at 8dc9329f (Linux, Swift 6.1.2): the temporary replay harness over the production TasteResidualEngine (byte-for-byte sources + swift-crypto shim) reports ALL FIXTURES REPLAY EXACTLY for all three golden household fixtures against the shipped 762-recipe catalog — confirming the merged engine (contract ingest + class-aware staleness) still reproduces the committed fixtures; the fixture files are untouched. Hosted iOS CI run 38844186301 runs on 8dc9329f.

How I checked it

All commands run in the sandbox at tree state d5ad8965 (the merge of the parallel alignment head into this branch; the round's re-runs are listed in the final bullet below):

• taste-store-check (Linux, Swift 6.1.2): 50/50 passed (./refresh.sh mirror sync + suite run; shared slice now 15 tests).

• community-check: 74/74 passed (bun source pins verified; Swift test linked with -Xlinker --allow-shlib-undefined — the swift::threading::fatal link gap is a Linux toolchain quirk, not a test failure).

• Substitution checker: OK — 70 swaps, 126 mapped endpoints, 329 composition keys, sweep pass (check_substitutions.sh over the canonical-key table).

• Python suites (pytest): substitutions + everyday tools 64 passed; data/recipes/tools 84 passed, 1 skipped (pre-existing skip).

• evaluation/taste-lab (python3 -m pytest tests): 10/10 passed after the engine alignment + fixture regen.

• Contracts (docs/spec/taste/contracts): check.py — all examples valid and self-consistent; mutation_matrix.py — all mutations caught, positive control passes.

• backend/community: tsc --noEmit clean; bun test 138 pass / 26 skip / 0 fail (after bun install; without it the Bun type defs break typecheck — environment, not code).

• backend/gemini-agent: tsc --noEmit clean; bun test src/__tests__ 338 pass / 0 fail; lechef suite (bun src/run/lechef-suite.ts) 65/65 job successes, 0 unsafe answers, 0 unauthorized actions, all gates true.

• evaluation/catalog-reach: python3 reach_report.py --repo-root ../.. --out ... --md ... reproduced the committed numbers after regen (86.0% fully resolvable).

• Re-run on the merged head d5ad8965: evaluation/taste-lab pytest 10/10 passed and python3 -m taste_lab validate-fixtures reports All fixtures replay exactly; community-check verification script clean and swift test 74/74 on the committed mirror refresh (same invocation flags as before).

• Re-run at 65928425 (Linux, Swift 6.4, no Xcode): a scratch package mirroring the production sources byte-for-byte (TasteResidualEngine, TasteEngineModels, TasteFeatureExtractorV2, TasteFixture, RecipeBundleV2; CryptoKit→swift-crypto shim, the taste-store-check pattern) replays the three golden household fixtures through the TasteRankerTests comparison loop — ranking order, per-feature contributions, learnedContribution, totalScore, profile and hidden statements — 7/7 tests passed (swift test, incl. testGoldenFixturesReplayExactly); evaluation/taste-lab pytest 10/10 passed; docs/spec/taste/contracts/check.py 19/19 examples valid; taste-store-check/Scripts/refresh.sh mirror in sync at the new engine pin (0b4a34b3…).

What I couldn't check

• No Xcode in this sandbox: the iOS app surfaces compile and run their XCTests only on the repo's hosted macOS CI, which runs on this PR's push. Linux swift test in FLFeatureLogic has a pre-existing no such module CoreGraphics failure, which is why the portable mirrors exist.
• reverse-meal-eval was not run (documented as unrun in the prior rounds of this join).
• The two repair commits I pushed before opening this PR (5a82b961, edd5cc8d) are authored as Obvious <obvious@obvious.ai> without the Co-authored-by: Sam Gu trailer — adding it would require rewriting pushed history (force-push), which is not allowed; attribution is recorded here.

Assumptions

• reach-after.json|md are generated artifacts that track the merged catalog state, so refreshing them on the join is correct; reach-before stays pinned to the pre-join baseline.

• The engine keeps a stray-late-rating correction path even though the contracts checker raises on re-rating an already-rated cook (in the app, corrections arrive as correction_applied events, which the checker ignores); the golden fixture does not exercise that path.

• .github is kept identical to the trunk branch (f6f33e2a) so this PR's CI matches the trunk's hosted macOS workflow.

• The PR targets main because the branch was cut from the main tip; no earlier PR ever existed for obv/burst8-fridgeluck-join (open or closed), so this is the join's first PR.

• Round result: CI run 38845246311 on 34d74e12 compiled cleanly (8m36s — past every earlier compile failure) and ran the suite with zero GRDB reentrant fatals — the snapshot reentrancy fix holds. Three LeChefPlanWeeklyTests tests then failed with no such table: cook_plan_line: the hand-built makeDatabase() fixture never creates the production cook-plan tables, which the now-working snapshot read touches. The parallel thread repaired the fixture standalone as fb0f868b (production CookSessionMigrations.migrate in the fixture — the same repair its PR test(ios): wire the overhaul into the app and prove the fridge-to-meal-to-taste path #182 carries for the GoalPath slice); that is the current branch head, with run 38846607091 in progress. A local goal-path merge that duplicated the fixture fix was discarded in favor of the shared minimal repair; PR test(ios): wire the overhaul into the app and prove the fridge-to-meal-to-taste path #182 lands through its own gate.

• Queue-boundary caveat (recorded as agreed): the reentrancy refactor is syntax-checked locally (swiftc -parse) but typed and tested only on hosted macOS CI — -parse does not catch missing-argument errors, which is exactly how two bridged: arguments slipped through to run 38844186301 and were fixed in ead103e1/7456f537.

• Round result: CI run 38846607091 on fb0f868b fixed the missing-table failures but exposed one stale expectation: testCatalogPantryFixtureYieldsAFeasibleWeek pinned the pre-substitution shopping list ["2:100.0", "3:200.0", "4:200.0"], while the snapshot now legitimately carries the shipped reviewed egg→tofu binder swap (substitutions-v2.json egg-binder-to-silken-tofu, ratio 1.0) bridged to legacy ids 2→3 — the planner covers egg's 100 g shortfall with tofu, yielding ["3:300.0", "4:200.0"]. That swap was the first to bridge cleanly because egg and tofu both resolve in the fixture; the pre-reentrancy-fix trees never reached this path (fatal) or lacked the bridge rows. The parallel thread repaired the test as 1c22f76b — updated list plus a positive assertion that the reviewed 2→3 swap actually rides the dinner lines — and run 38847597335 is in progress on that head. Diagnosis independently reproduced from the shipped swap JSON and fixture bridge before adopting the shared fix.

• Round result: CI run 38847597335 on 1c22f76b is fully green — build-and-test (5m40s) and dead-code-advisory (3m55s) both pass; 666 tests executed, 1 skipped, 0 failures, zero GRDB reentrant fatals across the whole suite. The join branch compiles the merged overhaul stack and passes every weekly-plan, catalog, cook-session, and GoalPath-adjacent suite on hosted macOS. PR [obvious-fleet] join: integrate the overhaul slices (burst5/6/8) into one branch from main #181 is marked ready for review on branch head 1c22f76b; merge is the human's (squash per repo policy).

🔗 Obvious Project · 🧵 Obvious Thread · 🔀 Obvious PR Artifact

🔗 Obvious Project · 🧵 Obvious Thread · 🔀 Obvious PR Artifact

🔗 Obvious Project · 🧵 Obvious Thread · 🔀 Obvious PR Artifact

CI round — weekly-plan fixture (fb0f868)

• Run 38845246311 on 34d74e1 compiled cleanly and ran the suite with zero GRDB reentrant fatals — the held-connection bridge repair held through a real test run. Three weekly-plan tests failed: SQLite error 1: no such table: cook_plan_line — the catalog adapter tests build their database through the hand-rolled makeDatabase() fixture, which never created the cook-plan tables that the snapshot load reads for plan reservations (PlanDataAdapter reservations query).
• Fix (fb0f868, test-only): the fixture now runs production CookSessionMigrations after the core schema, mirroring what a real app database carries. Verified swiftc -parse clean; the round-trip test's explicit migrate call stays and is idempotent.
• Taste-parity and GRDB verification: Linux-side suites green earlier (taste-lab 10/10, contracts 19/19, store mirror 50/50); hosted macOS test verdict pending on the next run.

CI round — swap-aware feasibility expectation (1c22f76)

• Run 38846607091 (fb0f868): zero GRDB reentrant fatals, zero compile errors, 15/16 weekly tests green. One remaining failure: testCatalogPantryFixtureYieldsAFeasibleWeek expected [2:100, 3:200, 4:200], got [3:300, 4:200].
• Root cause: the expectation predates swap semantics. The shipped reviewed v2 table carries egg-binder-to-silken-tofu (binder ratio 1.0, plausible verdict and recheck); the allocator covers a short required line only when the substitute covers the shortfall fully ("no half-swaps"), so the egg 100 g shortfall rides the swap into tofu — egg never reaches the list, tofu's buy grows by exactly the shortfall. Production verified correct against PlanAllocating.allocateDinner and the bundle JSON; the test expectation was stale.
• Fix (1c22f76, test-only): expectation updated to the post-swap buys, plus an explicit assertion that the egg line carries a reviewed egg→tofu swap. swiftc -parse clean.

CI round — green (1c22f76)

• Run 38847597335 on 1c22f76: build-and-test success. Executed 666 tests, with 0 failures (0 unexpected); Test Suite 'All tests' passed; zero Database methods are not reentrant fatals; the previously failing testCatalogPantryFixtureYieldsAFeasibleWeek passed with the swap-aware expectation.
• Status: ready for human review and merge (humanMerge policy — no autonomous merge).

Greptile review round — 5542992530 verification record

All six findings were verified against the source at 1c22f76. Every one is real. The PR merged (into alpha, merge commit 485ae82e) before any fix or test landed on the branch, so none of the six is fixed in this merge; the verdicts below come from code reads (file:line at 1c22f76), not from failing tests.

• Finding 1 — catalog guides open empty (CookingGuideView.swift:63): confirmed. Catalog results materialize Recipe(id: nil, instructions: "") (RecipeRepositoryCatalogRecommend.swift:138-148), and the guide resolves catalog content only through the legacy-id bridge (catalogRecipeID(forLegacyRecipeID:)), so Start Cooking on a catalog-native result opens a guide with zero steps and zero ingredients; RecipeResultsView.handleStartCooking passes the raw result through. A reviewed fix exists on burst9 (commit 770ab5eb: makeCatalogContent resolves the packet through scoredRecipe.catalog?.ref.id, with the legacy bridge as fallback, and steps/ingredients load from it) — inbound via PR feat(release): the release-candidate build with every late fix, the goal path and the safety sweep green #188.
• Finding 2 — catalog meals lose nutrition (MealLogService.swift:99): confirmed (journal, deductions, and a dead reader). For a catalog-native cook, resolvePersistedRecipeID inserts a header-only recipes row (no same-title legacy row to match), so computeConsumedMacros (journal) and consumptionRequests (inventory deductions) — both reading recipe_ingredients — see zero grams, and catalogNutrition(forHistoryID:) has no production caller. The frozen-snapshot journal read is fixed in 770ab5eb (PR feat(release): the release-candidate build with every late fix, the goal path and the safety sweep green #188). The deduction half — consuming the packet's bridged lines through applyConsumption — has no fix on any branch; it is feature-sized and left open.
• Finding 3 — timer alerts stop matching (CookModeViewModel.swift:305): confirmed; open on every branch. resumeTimer and adjustTimer pass label: nil; perform schedules only if let label; pausing emits the notification cancel, so a resumed timer re-arms nothing and an adjusted timer keeps the stale deadline. The label lives in the store record (CookTimerStore), so the fix is to pass it into perform. The view model imports UserNotifications, so a fix plus a scheduler-spy test validate on hosted macOS CI only, not the portable Linux harness.
• Finding 4 — dinner stock counted twice (PlanAllocating.swift:477-483): confirmed within one dinner. Each line reads the unchanged pool while a swap in the same dinner consumes from the same pool; reservation happens only after the whole dinner is picked, so a swap-consumed substitute and a direct line of one ingredient can both claim the same grams. Cross-dinner reservation is correct (the weekly fixture proves it). A reviewed fix exists on burst9 (a per-dinner claimed map every line reads against) — inbound via PR feat(release): the release-candidate build with every late fix, the goal path and the safety sweep green #188. A fails-first portable test (one dinner whose short required line swaps into an ingredient the same dinner also consumes directly; total consumed never exceeds pool stock) was designed but not committed here.
• Finding 5 — saved swaps reserve wrong stock (PlanDataAdapter.swift:280-282): confirmed; open on every branch. Swap-covered lines are saved under the original ingredient with the full requirement and no shortfall, and loadReservedGrams selects no swap column, so it reserves the full requirement against the original while the substitute that actually covered the shortfall reserves nothing and stays double-bookable. Fix shape: persist the consumed split on PlanSwap (optional originalGrams/substituteGrams, back-compatible through the wholesale swap JSON codec) and reserve the split on load, with the legacy math as the fallback for old rows.
• Finding 6 — catalog ignores keto (RecipeRepositoryCatalogRecommend.swift:94): confirmed. The diet gate handles vegetarian and vegan only; keto maps to the .lowCarb mask on the legacy pool while RecipeCatalogClaims has no low-carb field, so no catalog recipe can prove compliance yet all of them pass the gate. The burst9 fix fails the gate closed for keto (case "keto": continue) — inbound via PR feat(release): the release-candidate build with every late fix, the goal path and the safety sweep green #188.

Round outcome: six of six findings confirmed; zero fixed in this merge. Findings 1, 2, 4, and 6 land with PR #188 (burst9 rejoin, base alpha). Findings 3 and 5 have designed fix shapes but no fix on any branch. No commits, tests, or CI runs were produced for this round — nothing was pushed to the branch, so there is nothing to run them against.

[coderabbitai]

Important

Review skipped

Bot user detected.

To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: de0ea2a7-1f43-4957-8745-953dccfc964e

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[codemetrics-ai]

Note

Review skipped

Codemetrics is installed on this repository, but it hasn't been added to your team yet.

To enable AI code reviews for this repo, go to your dashboard and add FridgeLuck to your team's repositories.

[greptile-apps]

Confidence Score: 0/5

[High impact] Integrates multiple feature branches with broad app-layer changes.

Fix the catalog cooking and logging paths, timer alerts, substitute stock accounting, and keto filtering before merging.

Findings

1.  Catalog guides open empty ▶
2.  Catalog meals lose nutrition ▶
3.  Timer alerts stop matching ▶
4.  Dinner stock gets counted twice ▶
5.  Saved swaps reserve wrong stock ▶
6.  Catalog ignores keto choice ▶

Fix with agent prompt

### Issue 1
apps/ios/Feature/Recipe/CookingGuideView.swift:59-63
Catalog results have `recipe.id: nil` and empty `instructions`. This lookup requires a legacy integer ID, while the guide also reads its steps from those empty instructions and loads ingredients only through that ID. Selecting “Start Cooking” therefore opens an empty guide even though the catalog packet contains the recipe. Resolve the packet through `scoredRecipe.catalog.ref` and use its lines and steps.

### Issue 2
apps/ios/Platform/Persistence/Services/MealLogService.swift:96-99
When a catalog recipe has no same-title legacy recipe, `resolvePersistedRecipeID` inserts only a `recipes` header. It creates none of the `recipe_ingredients` rows used by inventory deductions and journal totals. Logging that meal therefore records zero nutrition and deducts no ingredients. The frozen catalog nutrition reader is not used by production reports. Add a catalog-aware path for logging, reports, and deductions.

### Issue 3
apps/ios/Feature/CookMode/CookModeViewModel.swift:302-305
Resuming or adjusting a timer passes `label: nil`, so this branch drops the notification update. Pausing already cancelled the notification, which means a resumed timer has no alert. Adjusting a running timer leaves its alert at the old time. Look up the stored timer label when applying the scheduling effect.

### Issue 4
apps/ios/FeatureLogic/CookMode/Plan/PlanAllocating.swift:477-483
`allocateDinner` reads the unchanged `pool` for every line and reserves stock only after choosing the whole dinner. A swap can therefore consume an ingredient that another line still counts as available. For example, a 100 g shortfall can claim 100 g of substitute stock, then the substitute’s own recipe line can claim those same grams. Both appear sufficient and the shopping list misses the shortage. Reserve each allocation in a local working pool before checking the next line.

### Issue 5
apps/ios/Platform/Persistence/KitchenPlan/PlanDataAdapter.swift:280-282
A swap-covered line is saved under the original ingredient with its full requirement and no shortfall. `loadReservedGrams` ignores the swap and reserves that full amount against the original. It leaves the substitute stock available for another plan, even though the saved dinner needs it. Persist the actual original and substitute amounts consumed, then restore reservations from those amounts.

### Issue 6
apps/ios/Platform/Persistence/Repository/RecipeRepositoryCatalogRecommend.swift:87-94
Legacy search requires the `.lowCarb` tag for a keto profile, but this catalog gate checks only vegetarian and vegan claims. A keto user with ingredients for a high-carb catalog recipe can therefore receive it alongside results that passed the existing diet filter. Apply an equivalent low-carb check, or exclude catalog recipes whose compatibility cannot be established.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

This integration joins the recipe catalog, taste learning, LeChef tools, community service, and supporting checks.

• Catalog recipes join pantry matches and cooking flows.
• Onboarding adds only ingredients found in kitchen photos.
• The planner builds dinner plans from available kitchen stock.
• Cook Mode guides recipes with saved steps and timers.
• Live cooking suggestions wait for approval on the phone.
• Cooking and recipe choices shape a local taste profile.
• Community features connect through typed app and server paths.
• Reverse scan previews and logs the chosen recipe consistently.

Acknowledged limitations:

• obvious-autobuild[bot] records the residual taste ranker as intentionally off by default.
• obvious-autobuild[bot] records the late-rating correction path as intentional.
• obvious-autobuild[bot] records reverse-meal-eval as unrun.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  Catalog[Catalog recipe] --> Match[Catalog recommendation]
  Match --> Preview[Preview reads catalog packet]
  Match --> Guide[Guide expects legacy recipe ID]
  Guide --> Empty[Catalog-only guide has no steps]
  Match --> Log[Meal logger resolves integer ID]
  Log --> Header[Creates recipe header only]
  Header --> Reports[Reports and deductions read legacy ingredients]
  Reports --> Zero[No ingredient rows means zero totals]

Loading

_{Reviews (1) · Last reviewed commit: "fix(ios): expect the reviewed egg-tofu s..." · Reviewed by Greptile}