test: A4 compaction phase 4 — plan-level test hardening (plan #951, source #552)

[btipling]

Plan

Implements plan #951 — HANDOFF-READY (source #552 A4 compaction engine, parent #947 phase 4). Test-only hardening: no product code, no caps changed, no Production mutate.

What shipped

Coverage map vs the parent testing matrix (Phases 1–3 already prove rows 2a–2h, confused-deputy, planted/stale, pin/clip, checkpoint writer, seam fail-closed). This PR closes the two composition-level gaps the plan locks:

1. Redis copy-forward composition (lib/sessions/sessionStore.test.ts) — plan #951 row: RedisSessionStore.upsertEnvelope copy-forwards compactionPointer from LWW existing when a host flatten omit arrives (durable, no volatility); a sibling modelMessagesPointer write does not disturb the carrier; read-back round-trips. Closes the gap where only the Memory store path was proven — Redis is what Production resolves through.
2. Workflow-boundary fail-open (lib/workflows/turnLoop.test.ts) — plan #951 row: summarize_failed on the compact path fail-opens; the turn completes from the un-compacted projection AND no fold anywhere in the run carries a fabricated compactionCheckpoint (a partial summary must never become the next turn's prefer-checkpoint seed).

Testing matrix (plan #951 rows 5–8 + row 3 re-proof)

Row	Status
Copy-forward (row 5)	✅ helper + Memory (#949) + new Redis composition row
Confused-deputy (row 6)	✅ existing route rows 2b/2e (unbound → no blob read; malformed → fallback, no 5xx)
Planted/stale checkpoint (row 8)	✅ existing rows 2b/2c/2h + adversarial #954 re-bound rows
Fail-open (row 7)	✅ existing loop rows + new no-fabricated-checkpoint row
Checkpoint-preferred seed (row 3 re-proof)	✅ existing row 2a [summaryRow, ...re-paired tail]

Gates (required, run in this cloud workspace)

• npm run typecheck — clean
• npm test — 177 files / 3527 tests green (di-gate + drizzle-journal-gate pass)
• npm run build — green (pre-existing jose/Edge-Runtime warnings only)
• Focused: vitest run lib/sessions/sessionStore.test.ts lib/workflows/turnLoop.test.ts — 222 tests green
• Environment note: npm install + npm run fetch-harness were required in the fresh sandbox checkout (stale public/harness/harness.wasm caused 13 local wasm-int failures; fresh artifact resolves all).

Not in scope

Living docs (Phase 5 / #952), caps (none changed), Production mutate (none).

Closes #951 (phase acceptance pending adversarial review; no merge in this step).

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated
invincible	Ignored		Sep 7, 2026 12:13am UTC

[btipling]

Adversarial review — PR #956

Verdict: PASS WITH NOTES
Repo: btipling/invincible
Scope: main ← plan/a4-compaction-p4-test-hardening (fdc62ca) · 2 files · test-only hardening (plan #951)
Lenses run: L1, L6 (skip: L2 product-security — test-only, no API/secret/runner reach; DI/cost gate ran: injected fakeClient, no PGlite/createDbConnection; L3 no UI/Wasm; L4 no CI/artifact; L5 no new runtime loop; L7 no config/host bind; L8 comments are load-bearing; L9 no UI)
AGENTS.md read: yes (tests-direct-with-vitest; no product-code / cap / docs in this phase)

Existing Phase-2/3 rows the PR body cites as already proving parent matrix 3/6/8 (route.test.ts 2a/2b/2e/2h, loop fail-open/reconstruct) are present on this head. Confused-deputy 2e is unbound-cp → legacy with no blob read; 2a is checkpoint-preferred seed with honesty + re-pair. This pass attacks the two new rows.

Findings

Sev	Lens	Finding	Break scenario	Refutation attempt	Confidence
Minor	L1+L6	RedisSessionStore.upsertEnvelope copy-forward row (lib/sessions/sessionStore.test.ts) locks host-omit + sibling mm + read-back, but not “intentional worker-authored value still wins” — the other half of the plan #951 / #949 copy-forward lock, which the Memory row immediately above does assert (t_cp_keep → omit keep → t_cp_new wins). Redis upsertEnvelope is a duplicated helper nest, not a shared function; wrap-order meta: { ...copyForward(input, existing), ...existing } keeps omit-copy-forward green and silently loses the next compaction’s pointer.	1. Production Redis store. 2. Worker writes compactionPointer: t_cp_s_0001. 3. Host flatten omits the key (already asserted). 4. Next compact persist PUTs compactionPointer: t_cp_s_0002. 5. If Redis applies stored-meta last, envelope stays t_cp_s_0001. 6. Next turn prefer-checkpoint seeds the stale Blob. This test still passes.	“Memory + copyForwardCompactionPointer unit already prove explicit-wins.” Those do not execute redisSessionStore.ts. This PR exists because Memory-only left a wrap-order hole on the store Production resolves. “Identical helper nest today.” Today, yes — the row does not lock it, so a Redis-only edit is free to invert precedence.	high
Nit	L6	New summarize_failed fail-open row (lib/workflows/turnLoop.test.ts) asserts ckFolds length 0. That is true if persist is never called with a fold at all ([].filter is empty). The wall_clock row on this file already does expect(foldCall).toBeDefined() before the undefined-checkpoint check. status === 'completed' is a proxy, not a fold lock.	A future persist short-circuit that skips fold on fail-open complete still reports completed + ckFolds.length === 0. The “no fabricated checkpoint” claim is then unproven.	“wiredDeps persist always runs on complete.” Today it does. The assertion still does not require a fold object, unlike the wall_clock sibling.	medium

Residual risk

Prefix-clip / reconstruct fail-open paths already lock compactionCheckpoint undefined on those seeds; this new row covers summarize_failed + priorMessages (the test-only seed). Production compact omits priorMessages — compactedCheckpoint is a single writer, so the fold contract still holds, but this row does not re-prove reconstruct-path folds. Unbound compactionPointer + bound modelMessagesPointer composition is implied by 2e+2d, not a dedicated row. No live Redis/Workflows/Gateway in this review.

Merge guidance

• PASS WITH NOTES: safe to merge from this attack; the Minor is cheap and should land on this PR (Redis explicit-wins + read-back).
• Do not change any existing cap. Do not add product code. Phase 5 / #952 still owns living docs.
• Redis row: after host-omit / sibling mm, a later upsertEnvelope with compactionPointer: t_cp_s_0002 must win, modelMessagesPointer must copy-forward, readEnvelope must round-trip the new id.
• Fail-open row (optional Nit): expect(persistSpy.mock.calls.some((c) => c[0].fold !== undefined)).toBe(true) before the no-checkpoint filter.

What was not attacked

Live Redis, prod Gateway, Workflows step isolate, DO runner, Wasm/DOM paint, npm run build in this workspace, Phase 5 docs, whether fakeClient JSON-round-trips the real node-redis adapter (pre-existing seam).

[btipling]

Adversarial #956 follow-up 1 (landed on this PR)

Posted the PASS WITH NOTES review (on fdc62ca), then implemented the surviving Minor + Nit on 8ad0231:

1. Redis explicit-wins — after host-omit + sibling mm write, a worker compactionPointer: t_cp_s_0002 wins and modelMessagesPointer copy-forwards; readEnvelope round-trips both. Omit-only could not catch stored-meta-last wrap-order on the duplicated Redis nest.

2. Fail-open fold existence — summarize_failed now requires a persist fold before the no-compactionCheckpoint filter (same vacuity lock as the wall_clock sibling).

Focused: vitest run lib/sessions/sessionStore.test.ts lib/workflows/turnLoop.test.ts — 222 tests green.

No product code. No cap change. Refs #951. Refs #947.