Wave 46 dim 19 ACLM booster PR — STATE¶

Date: 2026-05-12 Branch: feat/jpcite_2026_05_12_wave46_dim19_ACLM_booster Worktree: /tmp/jpcite-w46-dim19-ACLM Tick: Wave 46 永遠ループ tick6 #7 Author: parallel subagent (loop) Lane claim: /tmp/jpcite-w46-dim19-ACLM.lane

Scope — 4 dims, 1 booster PR¶

Closes 3 distinct dim 19 sub-criterion gaps and explicitly documents the 4th as redundant (already landed). Memory anchors: feedback_dual_cli_lane_atomic / feedback_completion_gate_minimal / feedback_session_context_design / feedback_rule_tree_branching.

dim	code	action	files	LOC
A	semantic_search legacy v1 MCP wrap	NEW MCP tool wrapping `POST /v1/semantic_search` (1024-dim e5-large canonical vec corpus)	`src/jpintel_mcp/mcp/autonomath_tools/semantic_search_legacy.py` + `__init__.py` registration	~145
C	amendment_diff ETL test extension	NEW test module covering canonical_value sort, Japanese unicode equivalence, projection_regression_candidate logic, dry-run no-op, hash-parity gating	`tests/test_dim_c_amendment_diff_etl_v3.py`	~228
L	session_context (state token 24h TTL)	NEW REST router with `POST /v1/session/{open,step,close}` + in-process OrderedDict LRU + 24h TTL + 32-step cap + 16 KiB saved_context cap; wired into `main.py` via `_include_experimental_router`	`src/jpintel_mcp/api/session_context.py` + `main.py` wire	~270
L	session_context tests	NEW pytest module covering open/step/close happy path, 410 on unknown/expired, 413 on step cap and oversize saved_context, 422 on token shape, store_stats introspection	`tests/test_dimension_l_session_context.py`	~175
M	rule_tree_branching	SKIPPED — `src/jpintel_mcp/api/rule_tree_eval.py` (412 LOC) + `tests/test_dimension_k_rule_tree.py` (387 LOC) already landed in PR #139 (Wave 46 dim 19 KN booster). dim M = "multi-step rule tree" is conceptually the same surface (AND/OR/XOR tree eval) — extending here would duplicate `_evaluate_node`. Logged here for traceability per `feedback_completion_gate_minimal`.	(none)	0

Architecture notes¶

Dim A (legacy v1 MCP wrap)¶

Wraps the older 1024-dim canonical vec corpus surface (api/semantic_search.py, migration 166). The modern hybrid v2 surface (384-dim e5-small + cross-encoder reranker) already has an MCP wrap at mcp/autonomath_tools/semantic_search_v2.py. The legacy surface had no MCP wrap → agents with e5-large embeddings had to HTTP-hop. Closed here.
Single ¥3/req billing event (parity with REST handler _billing_unit = 1).
_ENABLED flag via JPCITE_SEMANTIC_SEARCH_LEGACY_ENABLED (default ON). __init__.py registers via decorator side-effect.
NO LLM API — client supplies the 1024-dim L2-normalised embedding.

Dim C (ETL test extension)¶

9 new test cases targeting scripts/etl/backfill_amendment_diff_from_snapshots.py.
Covers:
canonical_value sort stability on target_set_json,
empty equivalence collapse (None / "" / "[]"),
should_record_field_change plain and unicode-Japanese path,
projection_regression_candidate fires iff raw byte drift + hash parity (eligibility/summary/raw_snapshot all match),
dry-run (apply=False) NEVER touches am_amendment_diff,
apply→dry-run idempotency metrics (candidate_diffs_new == 0).
In-memory sqlite, no autonomath.db dependency, no network, no LLM.

Dim L (session_context, NEW)¶

3 endpoints under /v1/session:
POST /open — returns state_token (32 hex chars) + 24h expires_at + initial saved_context.
POST /step — appends a turn under existing token; LRU-bumps to end; returns cumulative count + last_step_at.
POST /close — removes the token + returns final snapshot + step_log. Second close → 410.
In-process OrderedDict + threading.Lock (no Redis, no sqlite — per feedback_zero_touch_solo + feedback_no_quick_check_on_huge_sqlite).
Caps: _MAX_SESSIONS=10_000 (LRU evict), SESSION_TTL_SEC=86_400 (24h), 16 KiB saved_context, 4 KiB step, 32 steps.
Honest gap: in-process means session does NOT survive Fly machine restart or autoscale to a different machine. Disclaimer surfaces this. 99% of agent multi-turn loops fit in a single machine + 24h.
§52 / §47条の2 / §72 / §1 disclaimer on every response (verified by test).

Dim M¶

Skipped with explicit rationale (see table above). Avoids violating feedback_completion_gate_minimal ("最低 blocker に絞れ" — don't invent surface just to fill a slot).

Verify¶

Inside the worktree (/tmp/jpcite-w46-dim19-ACLM):

# ruff (4 new files)
.venv/bin/ruff check \
  src/jpintel_mcp/api/session_context.py \
  src/jpintel_mcp/mcp/autonomath_tools/semantic_search_legacy.py \
  tests/test_dimension_l_session_context.py \
  tests/test_dim_c_amendment_diff_etl_v3.py
# All checks passed!

# pytest (3 pre-existing + 20 new)
.venv/bin/python -m pytest \
  tests/test_a3_amendment_diff_etl.py \
  tests/test_dimension_l_session_context.py \
  tests/test_dim_c_amendment_diff_etl_v3.py \
  -q
# 23 passed, 2 warnings in 1.16s

Constraints honored¶

feedback_dual_cli_lane_atomic — lane via mkdir /tmp/jpcite-w46-dim19-ACLM.lane, worktree on origin/main.
feedback_destruction_free_organization — no rm/mv, only adds + 1 wire in main.py + 1 import in autonomath_tools/__init__.py.
feedback_autonomath_no_api_use / feedback_no_operator_llm_api — zero LLM SDK imports; banned-import regex in dim L test fails the CI if regressed.
feedback_overwrite_stale_state — historical STATE docs left as-is; this doc is SOT for the ACLM booster.

Outcome target¶

dim 19 audit re-score should lift: * Dim A mcp_wrap_count (legacy v1 wrap now exists) → +1 sub-criterion. * Dim C tests_present (new dedicated v3 test) → +1 sub-criterion. * Dim L rest_surface_present + tests_present (new fully) → +2 sub-criteria. * Dim M unchanged (already covered by PR #139 rule_tree_eval).