feat(home-a): conversation-first parent home prototype (synthetic data)

[obvious-autobuild]

Candidate A of the parent-home A/B comparison — DO NOT AUTO-MERGE.
Human author: Gilbert Polanco (gilbertpolanco42@gmail.com)

• Evidence artifact: Parent Home Candidate A — Conversation-First Prototype: Evidence & Findings (art_6oQvZxHx)
• Task: todo_CDoPnsFZ (parent home candidate A — conversation-first)
• Live preview (E2B session URL, expires with the sandbox session): https://8081-ibr1ped9y83529wciyqg1.e2b.app
• Synthetic data only. No real user data anywhere; the Polanco household is fictional.

Acceptance criteria

1. Runnable Expo RN prototype implementing the conversation-first objective: the home IS a conversation — capture, in-thread review of interpreted events, publish with raw-preserved provenance, conversational append-only correction, catch-up digest + month view, and cited read-only questions in one thread surface.
2. Voice-primary capture with a full text alternative hitting the SAME pipeline (raw transcript → extraction → proposed events); real device voice is not wired (no microphone in this environment) — the text path is fully functional and voice ships as a simulated-transcription demo mode, disclosed in the evidence doc.
3. Proposed events (type, child, time, amount, audience) are inspectable and editable in-thread before publish; publish confirms in-thread with raw source preserved.
4. Post-publish correction is conversational and appends lineage; the original stays intact.
5. Catch-up digest ("what happened since you last checked in"), month view on request, and read-only questions answered from the synthetic history with source citations; answers never write (asserted via store mutation counter).
6. Vitest coverage: mixed-topic capture through BOTH text and simulated-voice paths; proposal edit before publish; publish → raw-preserved entry + events with provenance; correction appends lineage; read-only answers never write; fail-closed authorization (non-member/anonymous sees nothing).
7. Repo gates green: pnpm turbo run typecheck && pnpm turbo run test (plus the app's Vitest suite).

Why

The parent-home A/B comparison needs two equally complete candidates to test whether a conversation surface beats a feed-first home for capture→review→correction→catch-up. Candidate A builds the conversation-first thesis end-to-end so the comparison is between real implementations, not stubs.

What changed

New self-contained workspace app apps/home-a (@journal/home-a) — no shared code touched:

• src/journal/store.ts — in-memory store mirroring Convex adapter semantics: schema-decoded entries/events through packages/domain Effect v4 schemas, raw-before-events capture, captureId replay protection, stale-extraction suppression, reviewable proposals, publish receipts with provenance, append-only corrections, audience-filtered reads, citations, read-only questions, catch-up/month views, mutation counter for no-write assertions.
• src/conversation/controller.ts — unified text/simulated-voice routing, question detection, review/publish/correct/catch-up/month-view, per-member principal switching (publish/correct are arrow class fields so the UI can pass them as bare references).
• src/capture/extraction.ts — deterministic rule-based extraction double (no LLM): mixed-topic clause splitting, per-child attribution, six canonical categories + school transition, quantity/time parsing, unknown content (e.g. health) stays raw, v0.3 canonical ExtractionRequest/ExtractionResult envelope with producedBy lineage.
• src/fixtures/ — synthetic Polanco household + September 1–16, 2026 dataset (24 captures): Sep 2 pre-K start, Sep 5 park photo note, Sep 8 balance-bike milestone, Nana Rosa's Sep 3/10/15 notes, a parents-only entry (Sep 6), the Sep 12 nap 30→60 lineage example, and the two required mixed-topic utterances.
• src/ui/ + app shell — sign-in (Dana / Gilbert / Rosa / anonymous fail-closed), conversation surface, editable review cards, receipts with correction affordance, answer citations.

Tradeoffs

• Self-contained app over shared-shell integration: keeps the candidate isolated for A/B (no changes to apps/mobile, backend, domain, CI), at the price of some duplicated UI primitives; consolidation is a post-selection follow-up.
• Simulated voice over device voice: no microphone exists in this environment; the transcriber seam matches the voice spike's surface so a real transcriber swaps in without UI changes.
• In-memory store over live Convex: preserves raw-before-events + captureId idempotency + schema-validated data without a backend dependency (backend thin-path port is another owner's lane).

Shared-change requests

None blocking. If Candidate A wins, follow-ups that would touch shared files: wire apps/home-a into CI's affected graph; native voice transcriber behind a device gate.

Verification

All commands run against exact HEAD 0089db9483840ca1881c51a9ced522acea96949c (post-rebase):

• pnpm install --frozen-lockfile — clean
• pnpm turbo run typecheck — 11 tasks OK
• pnpm turbo run test — 8 tasks OK
• cd apps/home-a && npx vitest run — 2 files, 25/25 passed

Browser walkthrough (Expo web, real Chromium): sign-in ×4 profiles, mixed-topic voice capture → 3 attributed review cards, read-only naps answer with 9 cited records, publish receipt, correction before 50 min → after 60 min, catch-up digest with citations, month view, anonymous fail-closed (empty thread + denied turn), Rosa's digest excluding the parents-only Sep 6 entry. Screenshots: project files home-a tc-1..tc-11 (details in the evidence artifact). The walkthrough also caught and fixed a real defect unit tests missed: bare method references lost this (reading 'now' TypeError) — fixed in 0089db9.

🔗 Obvious Project · 🧵 Obvious Thread