quilt kernel P6: tidepool — the helper-thread memory ocean joins the ledger

[SuperInstance]

Why

The tidepool bench (sibling of hermit, same quilt kernel) proves the helper-thread memory ocean: every write is a dual-write — the ocean markdown is the human-readable memory, the quilt WAL is the replayable ledger. The bench contract lives in /tmp/tidepool/README.md; this PR ports it into hermit as a first-class module.

Import, never duplicate: the port consumes commitProjection from src/quilt/commit.ts and the hash-chain primitives (verifyChain, WalRow) from src/quilt/projection.ts. The fnv1a chain input is byte-identical to hermit's, so tidepool WAL rows share the same quilt_wal table without collision.

What

src/tidepool/ocean.ts — TidepoolOcean: markdown ocean + recall index. Deterministic weighted token scoring (helper/question ×2, response/meta ×1), ties to most recently remembered, limit clamped 3..5. stats() audits the rememberBotAction oaths: never modify (no edit/delete API exists), always append-only (re-remember appends a new section; a memory of a memory stays), always timestamps (every section carries rememberedAt).

src/tidepool/stall.ts — threadLengthMonitor as a pure function: detectQuietThreads (windowed flat-run detection, growth resets, unsorted input) and newlyQuietThreads (crossing-only delta for a 30-min cron — silence is remembered once; bench extra, pure composition).

src/tidepool/projection.ts — projectHelperThread BIND-for-BIND with the L1 law honored (channel + helper endpoints bound before the thread links them); projectQuietThread mirrors the P2 .refused negative ledger — a quiet thread is an absence, and the absence is the information. replayThreadIndexFromWal folds WAL-only rows; .quiet state survives alongside remembered threads. readWalRowsByMutation / readWalRowsSince tolerate both SqliteD1 (sync .all()) and Cloudflare D1 (async {results}) shapes.

src/tidepool/index.ts — the six public entry points:
rememberHelperThread, rememberChannelThreads (one mutation for N threads), recallHelperContext, rememberBotAction, drainQuietThreads, plus replayThreadIndexFromWal from projection. Every write dual-writes; WAL failures are caught and counted, never thrown into the helper path.

Migration 0015 — helper_threads + helper_logs (drizzle-generated, snapshot + journal committed; second generate run confirms "No schema changes"). Per the contract, the WAL is the source of truth — these tables are the materialized view for live reads.

Verification

• bun run typecheck clean
• New suite: 11 tests, 55 expects, all pass
• Full suite: 400 pass / 4 fail — the 4 fails are the pre-existing lobster-artwork failures on base (untouched here); p4 baseline was 393 + 11 = 404 total

Deviations from the README (documented per contract)

1. Migration number 0015, not 0014 — 0014 was already taken by the p4 quilt_wal_lock migration on this base.
2. replayThreadIndexFromWal is exported from projection.ts, not index.ts — keeps index.ts to the six write/read entry points; same public surface overall.
3. WAL failure counting uses a local counter in index.ts (getTidepoolWalFailures) — PR quilt kernel P5-ops: the ledger reports its own health #5's ops counters (recordWalFailure("tidepool_projection")) are a 2-line follow-up once quilt kernel P5-ops: the ledger reports its own health #5 lands; this branch is independent of quilt kernel P5-ops: the ledger reports its own health #5 by design.

Merge order

Stacked on quilt-kernel-p4-review-fixes (same base as PR #5). Independent of PR #5 — they touch disjoint files (P5: src/quilt/ops.ts + catch sites in data layer; P6: all-new src/tidepool/*). Both land after #4.

Refs openclaw#48