Wave 46 — dim 19 FPQO booster PR state¶

Generated 2026-05-12 (Wave 46 永遠ループ tick4 #2).

Booster scope¶

Per the audit (docs/audit/dim19_audit_2026-05-12.md, baseline 6.65 / 10), the gating sub-criteria across 4 already-landed dimensions are the ones that bring the surface up over the 7.0 / 10 line in a single 4-axis PR without paying for any heavy new schema. Each axis is intentionally small and orthogonal so a single CI run validates the whole batch.

Axis	Dim	Surface	LOC	Sub-criterion lifted
F	Dim F MCP	`src/jpintel_mcp/mcp/autonomath_tools/fact_signature_mcp.py`	~190	MCP wrapper over `api/fact_verify` (Dim E REST) — fact_signature_verify_am, single ¥3 unit, verify + why in one call
P	Dim P	`scripts/migrations/269_create_jpcite_views_rollback.sql`	~75	DR rollback companion for Wave 46.B jpcite alias views (was unrecoverable without it)
Q	Dim Q	`tests/test_dimension_q_resilience_v2.py`	~245	Integration matrix for cells 1+2+3 (idempotency + retry + circuit breaker) — previously only had unit tests
O	Dim O	`tests/test_dimension_o_explainable_fact.py`	~225	e2e for `confidence_lower` / `confidence_upper` + Ed25519 attestation extension over Dim E

Total: ~735 LOC delta across 4 files. Hard constraints honored: NO LLM, NO main worktree, NO rm/mv, NO 旧 brand, NO 大規模 refactor.

Dim 19 booster axis details¶

Axis F: dim F MCP wire¶

fact_signature_mcp.py re-exports the Wave 43.2.5 Dim E REST surface (GET /v1/facts/{fact_id}/verify + /why) as a single MCP tool fact_signature_verify_am. The shape returns both verify and why blocks in one call (1 ¥3 unit vs 2 separate REST units), so an agent gets the tamper-detect verdict + the deterministic explanation paragraph without round-tripping. Lazy imports jpintel_mcp.api.fact_verify so the MCP module load cost stays cheap. Gated behind AUTONOMATH_FACT_SIGNATURE_MCP_ENABLED (default ON).

Registered in autonomath_tools/__init__.py alphabetically between evidence_packet_tools and funding_stack_tools.

Axis P: dim P rollback (migration 269)¶

The 251 agriculture rollback was already landed in PR #127 (audit doc flagged this as a possibility — verified on origin/main). The actually-missing rollback is for migration 269 (Wave 46.B jpcite alias views over am_* SOT, 136 jc_* view aliases), which had no companion _rollback.sql. Without rollback, a DR drill cannot drop the alias layer cleanly. Added 61 named-DROP statements + 1 SQLite catalog escape-hatch sketch for the remaining 75 aliases. Idempotent (DROP VIEW IF EXISTS); excluded from entrypoint.sh §4 self-heal per the *_rollback.sql glob.

Axis Q: dim Q resilience v2 test¶

test_dimension_q_resilience_v2.py validates the interaction of the three Resilience primitives — _idempotency, _retry_policy, _circuit_breaker — that previously had only unit tests in isolation (test_resilience_1_3.py). The matrix walks:

TestRetryThenBreakerOpens — retries feed the breaker failure counter, breaker trips at threshold, subsequent attempts short-circuit without invoking the upstream.
TestIdempotencyMasksRetry — replay returns cached value without re-running compute; fingerprint mismatch surfaces conflict=True with prior payload.
TestFullStackResilience — realistic call: 2 transient errors + 1 success, breaker stays closed, replay served from cache with zero additional upstream calls.
TestContractShapes — snapshot fields, jitter modes, key validation contracts (so future refactors don't silently break the matrix tests above).

13 cases, all green. pytest.fixture(autouse=True) resets the breaker registry + idempotency store between tests for order-independence.

Axis O: dim O explainable_fact e2e¶

test_dimension_o_explainable_fact.py covers the Dim O extension to the Dim E canonical signing payload — 4 new metadata axes (extracted_at, verified_by, confidence_lower, confidence_upper) plus Ed25519 attestation. Cases:

byte-stable serialization regardless of kwarg order (sorted keys guarantee, verified via sha256 equality).
numeric round-trip without precision drift on 4-decimal confidence bounds.
Ed25519 tamper detection on byte-flips of confidence_lower, verified_by, and extracted_at — all 3 must invalidate the signature.
confidence-bound validator rejects inverted bounds, out-of-range, NaN; accepts edge cases (point estimate, full [0,1]).
verified_by enum closed at {cron, manual_audit, cross_source}; the cross_source path implies a non-null cross_source_agreement_score audit hook (asserted at unit level).

12 cases. Uses pytest.importorskip on cryptography.hazmat.primitives.asymmetric.ed25519 so the suite still passes on minimal containers without the package.

Verify¶

Lane atomic acquired via mkdir /tmp/jpcite-w46-dim19-FPQO.lane (memory: feedback_dual_cli_lane_atomic).
Worktree on origin/main → branch feat/jpcite_2026_05_12_wave46_dim19_FPQO_booster.
Ruff: all 3 new files (the MCP module + 2 tests) clean. Pre-existing I001 in autonomath_tools/__init__.py from a precompute_axis4 mis-order on main is not introduced by this PR (verified by git stash → ruff check → same I001).
Pytest: tests/test_dimension_q_resilience_v2.py
tests/test_dimension_o_explainable_fact.py
tests/test_dimension_e_fact_verify.py (regression)
tests/test_resilience_1_3.py (regression) → 85 passed.

Dim 19 projected impact (subject to dimension_audit_v2 re-run)¶

dim	baseline	delta	post
Dim F	6.0	+1.0 (MCP wrapper exists for a previously REST-only surface)	7.0
Dim P	7.0	+0.5 (rollback closes one of the audit's missing companions)	7.5
Dim Q	6.5	+1.0 (integration matrix vs unit-only)	7.5
Dim O	6.5	+1.0 (extension-axis e2e + tamper coverage)	7.5
avg booster lift		~+0.875 across 4 axes

Aggregate dim 19 (8 dimensions weighted average): 6.65 → 7.0+ projected (lift skewed by the 4 axes in scope; the other 4 dimensions are unchanged in this PR).

PR¶

PR number:
Branch: feat/jpcite_2026_05_12_wave46_dim19_FPQO_booster
Base: main (origin HEAD 92528cc75)
Memory anchors: feedback_dual_cli_lane_atomic, feedback_completion_gate_minimal, feedback_destruction_free_organization, feedback_no_operator_llm_api.