v2.7.0

F1 StratLab v2.7.0

September's PITWALL, RAG, and simulation work is on main. This release makes
the no-LLM PLAN timeline state explicit, keeps the configured provider when an
API request omits provider, and refreshes the FIA retrieval corpus through
2026. The deterministic scorer and both five-lap horizons are unchanged after
the #715/#724 evidence review returned NO-GO.

Measured results

Measure	Before	After	Change and scope
Issue-backed runtime bug fixes	-	10	#233, #264, #320, #1192, #1200, #1203, #1207, #1211, #1251, and #322.
Dependency-import guard fixes	-	1	Tests exercise SetFit through the existing shim; production code did not change.
Issues in the v2.7.0 completion scope	-	33 closed	14 were closed after publication from the 111 open at the start of the release sweep; 19 were already closed. #1196 was closed earlier and is excluded.
Unwaived OSV findings on the full main scan	2	0	100% fewer; one High and one Moderate finding cleared without waivers.
Lusail 2025 data build	270.4 s	80.0 s	70.4% less time, FastF1 cache warm, same machine.
Las Vegas 2025 cache read	3.99 s	0.69 s	82.7% less time, best of three fresh processes.
Las Vegas 2025 resident memory	4,042 MB	321 MB	92.1% less, same cache-read comparison.
Six-race cache footprint	2.1 GB	1.5 GB	About 29% less; source totals are rounded to 0.1 GB, so exact byte-level reduction is unknown.
Serial pytest run time	18m17s (1,453 tests)	8m41s (1,459 tests)	52.5% shorter in the recorded runs on the same machine. The counts differ, so this is not an equal-suite benchmark.
Wrong-season RAG hits in top five	43/75	0/75	100% fewer wrong-season hits across 15 season-scoped queries.
RAG precision@5 vs fixed-window baseline	0.109	0.217	+100% on the same 35 queries.
RAG MRR vs fixed-window baseline	0.368	0.679	+84.6% on the same 35 queries.
RAG citation match vs fixed-window baseline	0.657	0.829	+26.1% on the same 35 queries.

These are separate measurements with different inputs. They are not a promise
of the same speedup on every machine. The RAG trace set remains n=1, so its
1/1 citation match is a path check, not a general answer-quality estimate.
The runtime-fix count excludes test-only fixes such as #1191 and
evaluation-only fixes such as #1213 and #1222.

Other changes

• PITWALL now says when deterministic mode has no future pit plan. It does not
  invent a stop lap or compound.
• The backend API no longer replaces the configured provider when a request
  omits provider. Explicit overrides still mutate process-wide configuration,
  while the cached client can retain an earlier provider across sequential or
  concurrent mixed-provider requests. Follow-up #1261 tracks request-scoped
  client isolation.
• RAG uses the race season for replay retrieval, supports an optional season in
  chat, records the index manifest, and grounds citations in the actual tool
  trace. The 2026 FIA corpus uses article-aware 512/64 chunks; the 35-query report compares it with the historical fixed-window baseline and rejects 1024/128.
• The full OSV scan on main passes after updating Jupyter Server 2.20.0 to
  2.21.1 and PyJWT 2.13.0 to 2.14.0. No waiver was added.
• scripts/ had no safe deletions in the September audit. Documentation now
  distinguishes model-training telemetry from the FIA retrieval corpus.

Known limits

• Explicit API provider overrides still change process-wide configuration,
  while the cached client can retain an earlier provider across sequential or
  concurrent mixed-provider requests. #1261 tracks request-scoped client
  isolation.
• The wake gate is limited to deterministic mode. The rich profile remains
  ungated pending an authorized cold-wake measurement (#1201).
• The end-to-end Safety Car citation failure in #826 remains open. Tool-trace
  grounding does not yet guarantee correct application of conditional rules.
• The article-aware 1024/128 candidate was rejected; production remains on 512/64.