[Jammy2211]

Summary

PR 6 of #442, the last PR of likelihood-lifecycle phase 0B: a coverage matrix and a parity-witness convention. Stacked on #450 (PR 5), which is stacked on #449 (PR 3b). Merge in order: #449 → #450 → this. This PR's own commits are 3f40ce1e, 1f3d336d and the merge 6f733073.

• scripts/misc/tooling/coverage_matrix.py writes dashboard/coverage.{json,md}. For likelihood × class × measurement it reports current, stale, superseded or missing, read from the v3 dashboard/catalogue.json and the §3.3 provisional table. A new lint --check step stops it drifting, and it also fails if a planned cell's class disagrees with §3.3.
  • Today: of 246 planned cells, 2 are current, 6 stale, 0 superseded and 238 missing. Of the missing, 32 have only a nearest-size row; under the human size ruling (Delaunay 1250, rectangular 35² = 1225) they count as missing. The other 206 have no record. Parity is missing for all 123 planned setups, because no witnesses exist yet.
  • The 40 current records on planned cells all join their cell (2 + 6 + 32). The other 86 current records are compile/memory measurements or light_fixed variants, which no planned cell covers.
  • Stale rule: reused from the Brain profiling conductor (_profiling.py). A current record is stale when its release_pin is behind the catalogue's newest pin (currently 2026.10.17.1) or has no pin.
• catalogue/provisional_designations.json is a machine copy of the epic report's §3.3 table, checked cell by cell, with notes on the readings that took judgment. A test keeps it in step with catalogue_lifecycle.py.
• Parity witness convention: one parity.json per cell, documented in results/README.md and validated by parity_witness.py --check. It holds log L, the reference log L, a tolerance, status match / diverged / not_assessed, the reference source and date, and provenance. The builder emits a measurement: parity v3 record with metric abs_delta_log_likelihood and values {abs_delta_log_likelihood, tolerance}, because Pulse v3 requires non-negative dimensionless values. The signed values go in a parity block on the record. A not_assessed witness produces no record but is listed in parity_witnesses. No parity values were made up: there are 0 parity.json files, and the tests use fixtures only.

Scripts Changed

• scripts/misc/tooling/coverage_matrix.py (new): coverage matrix and --check
• scripts/misc/tooling/parity_witness.py (new): parity.json v1 schema, validator and --check
• scripts/misc/tooling/build_catalogue.py: reads parity.json into v3 parity records (history copies become superseded) and adds parity_witnesses
• catalogue/provisional_designations.json (new): §3.3 copy
• .github/workflows/lint.yml: parity_witness and coverage_matrix --check steps
• Tests: new test_coverage_matrix.py and test_parity_witness.py; parity cases added to test_build_catalogue.py
• Docs: results/README.md ("Parity witness"), catalogue/README.md, scripts/misc/tooling/README.md; regenerated dashboard/catalogue.json and dashboard/coverage.*

Follow-ups (non-blocking review notes)

• coverage_matrix.py:221: once real parity witnesses exist, their current records will be counted in current_records_unplanned. Exclude measurement == "parity" there.
• The §3.3 reading exists twice, in the JSON and hard-coded in catalogue_lifecycle.py. A test keeps them in step; later, have catalogue_lifecycle read the JSON.

Test Plan

• ruff 0.17.2 check and format --check
• every lint --check step, including the two new ones and build_catalogue --check --validate-with ../PyAutoPulse (valid at @3)
• pytest scripts/misc/*/test: 1600 passed, 6 skipped, 1 failed. The failure is the known test_hazards_prior_exit, which is environment-only and passes in CI.
• smoke 9/9
• browser test and lychee (CI)

Validation checklist (--auto run — plan was not pre-approved)

• Effective level: supervised (header: supervised, cap: feature/large → supervised). Explicit human auto launch on 2026-10-24. Judge tier: merge is a human /prm, after feat: drop the old results layout — v3-keyed trends, writers into the mirrored tree (#442 PR 3b) #449 and refactor: group scripts/misc tests by folder, READMEs per group (#442 PR 5) #450.
• Plan: on the issue (feat: likelihood-lifecycle 0B — mirrored results tree + profiling-summary@3 catalogue #442 comment "Auto run … PR 3b, PR 5, PR 6"), written at start, unmodified since
• Gate: tests as above · smoke 9/9 · review: an independent Opus reviewer returned CLEAN. The witness is basis-cited, but the parity-to-Pulse path is shown only by fixtures, since no real witness exists. Claims checked: the §3.3 transcription, the 2/6/0/238 arithmetic, that superseded is reachable, the stale rule, and parity against the Pulse v3 contract. · Heart YELLOW within the launch acknowledgement (same reason list as feat: drop the old results layout — v3-keyed trends, writers into the mirrored tree (#442 PR 3b) #449).
• Human: plan sound in hindsight?
• Human: diff matches plan (no scope creep)?
• Human: merge, amend, or reject — then log the outcome

Refs #442

🤖 Generated with Claude Code

https://claude.ai/code/session_01SeDHwsR2dnwCk8Btfhej2L