PR-5: record the 32 layer verdicts of the grand catalog (Claude lane; Codex lane absent) - #71

…layer narrative)

tools/sota-convergence/lane_packets.py builds stripped per-layer packets (incumbent rationale withheld, adopted kept,
seeded candidate order, catalog-wide manifest join by repository slug) with SHA256SUMS; lane-return.schema.json and
lane-prompt.md are the shared lane contract; record_verdicts.py validates each lane return against the layer-verdict
schema v2 rules, seals it under evidence/artifacts/layer-verdicts-20260922/<lane>/<run_id>.json, derives
pin/component_id/recipe_ref/platform_status in code, computes agreement (same_winner/disagree/codex_absent/pending)
and writes the ledger rows; build_verdicts.py renders the per-layer narrative in the handbook's generated block;
codex_lane.py runs the Codex lane per packet (resumable, usage captured). 100 tests cover the tools on synthetic
fixtures. Handbook block, evidence manifest and explorer regenerated.

Co-Authored-By: Claude Fable 5.1 <[email protected]>

… generated indexes

Found by rehearsing the record step on a full 32-layer lane run: the dated quality comparison mirrors each
layer's v1 overturn_when, so a recorded verdict now carries its own verdict_overturn_when (schema v2 field;
the recorded-verdict marker rule moves to it) and no v1 field is rewritten. Lane citations carry line
anchors and commentary; rows keep the bare canonical repository path and the sealed lane return keeps the
full citation. Citations that resolve to no repository file or name a generated index
(manifests/evidence.json, docs/ecosystem/index.html, which made the explorer embed its own hash) are
counted in one open gap. Rehearsal: record --write/--check, build_verdicts, landscape, validate.py,
convergence and the explorer fixed point all pass on 32 recorded rows; python3 -m unittest Ran 1513 OK.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…strict lane schema, winner/alternative exclusion)

Findings from the Opus evidence review of fc478d1..53545eb, each with a regression test:
- a recorded row never lists its winner among its alternatives (adjudicated disagreement case); a row left with
  no indexed alternative stays pending_lanes with the gap; scripts/landscape.py rejects a winner repository
  among alternatives;
- record --check compares every recomputed sealed file with the file on disk, and scripts/landscape.py verifies
  lanes.<lane>.sealed_sha256 against the sealed file's bytes;
- lane returns are checked against lane-return.schema.json properties at every level, an https challenger
  repository, and a why_selected that names one of its cited evidence paths;
- the packet file must hash to its SHA256SUMS entry (a missing entry gets its own message); a leak refusal while
  sealing rejects that lane instead of aborting the run; a malformed adjudication is reported, not ignored;
- sources_read absolute paths are rewritten to repository-relative paths before sealing;
- codex_lane.py decodes the partial output of a timed-out attempt instead of discarding it (mutation-checked);
- a no_selection row renders as "no selection" in the handbook narrative.
The packet withholding limit (adopted flag and foundation decision `selection` values reveal the incumbent
default) is documented with the comparison that would test it. python3 -m unittest Ran 1582 OK (skipped=204).

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

… source paths, adjudication leak)

The independent verification of cf68d27 confirmed seven findings resolved and three partly resolved, and raised
three new minors, all fixed here with tests:
- why_selected must cite an evidence path, as the lane contract says: one of its own cited paths or any
  repository-style path in the text (a lane may leave winner_evidence_refs empty rather than guess);
- sources_read is rewritten to repository-relative form only under a checkout passed as --lane-repo-root
  (repeatable; the same values reproduce the sealed text under --check), so a same-named file from another
  checkout is never attributed to this repository; every other absolute path is redacted by sanitize_value;
- a leak marker in an adjudication file is reported and leaves the row pending instead of aborting the run;
- tests now cover a recorded row reduced to no indexed alternative, landscape.py rejecting a winner listed among
  its alternatives, and the no_selection narrative label.
python3 -m unittest Ran 1587 OK (skipped=204).

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

…bsent, gap named)

Claude lane run wf_bf9a4521-4a9 (32 packets from the refreshed manifest; per layer an Opus proposer, two Opus
refuters with evidence and challenger lenses, one revision) plus the one-layer re-run wf_f9391ce2-0a2 for
us-equities/security-supply-chain after its first return failed the single-repository challenger rule. Every
completed child resolved to claude-opus-5-5; the coordinator set model.name from that measured provenance.
record_verdicts.py --write with --lane-repo-root: 32 recorded, 0 rejections; each accepted return sealed under
evidence/artifacts/layer-verdicts-20260922/claude/ with its sha256 in lanes.claude.sealed_sha256;
lanes.agreement is codex_absent on every row (Codex account usage limit until 2026-09-28) and each row names
that gap. v1 fields unchanged; verdict conditions live in verdict_overturn_when. Handbook block, join, evidence
manifest (32 sealed files hash-listed) and explorer regenerated. validate.py, landscape, record --check,
build_verdicts --check, convergence --all-recorded and explorer --check pass; python3 -m unittest Ran 1587 OK.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>