Measure LLM round 2: recall, repeatability, and .github visibility

[joshft]

Closes the last open item from the second external review. Protocol: blinded claim inventories written before any extractor run, two real merged changes as cases, six live runs total.

• Precision held: 115/115 proposals accurate, zero hallucinated files or mechanisms.
• Recall (new axis): 11/12 and 9/11 — every core behavioral claim surfaced; misses concentrate in subtle negatives, deliberate design omissions, and documentation-meta claims.
• Repeatability (new axis): exact-text stability across four identical runs is zero — the model rewords every proposal, so content-hashed ids never collide — while semantic stability is effectively complete (all runs cover the same claim set). The README states the consequence honestly: LLM remainder rows are stable in meaning, not in id or wording.
• .github visibility fix (the one code change, measured): .github diff sections now reach the model, ordered after all non-hidden sections so they can never crowd shipped code out of the byte cap. Unit tests pin the ordering and the crowd-out case; live A/B on a real workflow-touching change shows the workflow claim going from structurally unreachable to minted exact from the workflow section.

What the receipt cannot check here: the live measurements ran against past merged diffs in detached worktrees; the receipt covers the extractor change and its tests.

🤖 Generated with Claude Code

https://claude.ai/code/session_015GeUG1gboWiZSnFyzQghyp

[github-actions]

correctful receipt

10 claims — ✅ 10 verified · ❌ 0 refuted · ⚠️ 0 unverified

Change: origin/master...HEAD (ea2a528..8fa8d18 · input:2b956763aa06) — 3 files

⚠️ Unverified remainder (0) — what nothing checked

Empty — every harvested claim reached a probe.

✅ Verified (10)

Tier	Claim	Statement
T1-assertion	TestExtractMintsValidatedProposals	Extract Mints Validated Proposals
T1-assertion	TestFenceWrappedOutputTolerated	Fence Wrapped Output Tolerated
T1-assertion	TestProseWrappedOutputFailsLoud	Prose Wrapped Output Fails Loud
T1-assertion	TestAPIErrorSurfaces	APIError Surfaces
T1-assertion	TestByteCapIsHardAndDisclosed	Byte Cap Is Hard And Disclosed
T1-assertion	TestContextCancellationPropagates	Context Cancellation Propagates
T1-assertion	TestSectionFileParsing	Section File Parsing
T1-assertion	TestDotDirSectionsNeverReachTheModel	Dot Dir Sections Never Reach The Model
T1-assertion	TestGithubSectionsIncludedAfterCode	Github Sections Included After Code
T1-assertion	TestGithubSectionCannotCrowdOutCode	Github Section Cannot Crowd Out Code

Harvest coverage: 3 files — 1 claimed · 2 scanned · 0 unread

_{schema 0.0.6 · exit gate: refuted claims block; the remainder informs, never fails}