PR garrytan#1016 (upstream-fix/google-recipe-max-batch-tokens)
also touches two regression tests that pinned google as the canary
"real provider with no cap declared". Without backporting them to
fork master, `bun test test/ai/` would 2-fail on this tree even
though google.ts has been capped here since af2a806.

- no-batch-cap-suppression.serial.test.ts: replaces the "STILL warns
  for google" test with a stronger invariant — NO first-party recipe
  warns, because every native/openai-compat recipe now declares
  either max_batch_tokens or no_batch_cap.
- adaptive-embed-batch.test.ts: inverts `contractMatch.length > 0`
  to `=== 0` and adds google to the "must be absent from warnings"
  assertion list (sibling to voyage + openai).

Validation: `bun test test/ai/` → 144 pass / 0 fail.

Also: TODO.md updates link PR-1016 + PR-1017 from the related
entries; PR-2 oauth-bootstrap entry flips to FILED with the same
patch metadata.

On upstream merge of either PR, the next sync auto-drops these
local edits via clean text merge.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

JARVIS-ARCHITECTURE.md §6.25 records the 2026-05-15 follow-up session:
diagnostic correction (the historical 117k stderr "fire" that wasn't),
two upstream PRs filed (garrytan#1016 google max_batch_tokens, garrytan#1017
oauth_clients forward-bootstrap), CJK keyword-only eval verdict,
overlap-matrix verdict, M2-B + M2-C verdicts, mechanical cleanup
(48 sync_failures ack'd, bun-test hang root-caused), and the new
[facts:absorb] sub-process DB-connection latent bug. 8 TODO items
closed, 1 new entry filed.

CLAUDE.md "Chinese-first knowledge base" rule tightened per the
CJK-eval verdict: the strict invariant is *compound CJK (4+ Han chars
without whitespace) requires vector*; English and 2-3 char standalone
CJK match fine on keyword. Operational guidance unchanged — the modal
operator query on this brain is a compound CJK phrase that depends on
vector being live.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

v0.37.0.0 upstream sync (11 versions, 12 commits, v0.35.6.0 → v0.37.0.0).

Conflicts (4) resolved:
- CLAUDE.md: keep fork (upstream content lives in docs/CLAUDE-UPSTREAM.md;
  selected new sections will be ported in Phase 7).
- llms-full.txt: take upstream then bun run build:llms in Phase 2.
- test/ai/adaptive-embed-batch.test.ts: take upstream — v0.36.1.1 garrytan#1083
  cherry-pick superseded our PR garrytan#1016 backport tests.
- test/ai/no-batch-cap-suppression.serial.test.ts: take upstream — same.

Auto-merged cleanly: VERSION (→ 0.37.0.0), CHANGELOG.md, README.md,
package.json, bun.lock, .gitignore, skills/RESOLVER.md (fork
## KOS-Jarvis extensions section preserved at L128, upstream
skillpack-harvest entry added at L62), skills/manifest.json (47 total,
fork 4 KOS-Jarvis entries preserved), src/core/pglite-engine.ts (WAL
fork patch survived at L207 — line shift from L200 due to upstream
import additions).

Fork-protected paths (skills/kos-jarvis/, server/, workers/,
scripts/launchd/) untouched — verified via diff master..HEAD.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Phase 1 conflict resolution took upstream versions of these two
tests assuming v0.36.1.1 garrytan#1083 cherry-pick was the equivalent of
fork PR garrytan#1016. Wrong: v0.36.1.1 went with the "warn only for
the configured embedding provider" filter path (gateway.ts),
while fork kept the "declare max_batch_tokens on google recipe"
path (google.ts +7 lines, retained through merge).

Net effect of both paths together:
- gateway.ts warning filter narrows to configured provider (upstream)
- google.ts has max_batch_tokens declared so won't warn anyway (fork)

Fork tests expect 0 warnings; upstream tests expect 1 (google when
google configured). Took fork view back — google.ts business value
(pre-split CJK batches via max_batch_tokens=20_000 + chars_per_token=2,
reduces reactive token-cap discovery overhead) justifies retaining
the fork patch.

Verification: bun test test/ai/ → 224 pass / 0 fail / 764 expect().

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

JARVIS-ARCHITECTURE.md §6.29:
- 11 versions overview (v0.35.6.0 → v0.37.0.0, 12 commits, 333 files)
- 4-conflict resolution log (test/ai/ reverted to fork after finding
  v0.36.1.1 garrytan#1083 went a different fix path from PR garrytan#1016)
- Schema v66 → v78 migration evidence (12 migrations auto-bootstrap)
- ZeroEntropy lock procedure (config set ze_switch_declined_at +
  ze_switch_prompt_shown=true since CLI has no --decline flag)
- Health score 70 → 80 evidence + new v0.36.x checks all PASS
- Smoke evidence (typecheck/test/curl/query Chinese+English/patrol)

CLAUDE.md: Latest sync story pointer updated to §6.29.

TODO.md:
- Top intro replaced with v0.37.0.0 sync context
- L824 PR garrytan#1016 verdict revised — NOT superseded by v0.36.1.1
- Added 5 new P1 entries (post-v0.37.0.0):
  - dream-cycle LLM spend observation (calibration phases)
  - embedding_columns registry declare (optional)
  - image ingestion roadmap decision
  - doctor --remediate vs kos-patrol evaluation pointer
  - typed-claim fence for OH transcripts
- Branch cleanup deferred to Lucien (worktree refs block delete)

llms-full.txt regenerated via bun run build:llms.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…12 commits, v0.35.6.0 → v0.37.0.0)

333 files / +52455 / -3317 LoC. 4 conflicts resolved (CLAUDE.md /
llms-full.txt / 2 test/ai/ files — last 2 reverted to fork view after
discovering v0.36.1.1 garrytan#1083 went a different fix path from PR garrytan#1016).

Schema v66 → v78 auto-bootstrap (12 migrations, zero manual ALTER).
WAL fork patch survived at src/core/pglite-engine.ts:207. ZeroEntropy
switch lock applied (ze_switch_declined_at + ze_switch_prompt_shown).
Doctor health_score 70 → 80; 3140 pages preserved; vector(1536) unchanged.

Smoke: bun test test/ai/ 224/224 green. kos.chenge.ink/health returns
0.37.0.0 + engine=postgres. kos-patrol kickstart exit 0, dashboard
written. Chinese + English query smoke top-hits in healthy band.

Fork dirs unchanged (10). Branch cleanup deferred (worktree refs).
PR garrytan#1016 verdict revised: NOT superseded — upstream filter path
co-exists with fork declare-max_batch_tokens path.

Full story: docs/JARVIS-ARCHITECTURE.md §6.29.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>