lane_packets --manifest-newcomers: carry a discovery refuted only by a missing vote (F-2W-3)

[seathatflowsinourveins]

Summary

This is roadmap move F-2W-3 from docs/decisions/2026-09-28-ecosystem-roadmap.md (#469).

The problem. In catalogs/sota-convergence/manifest-20260926.json, 27 discoveries are refuted_* only because a vote never returned: the GPT-6 fit vote is {missing: true} in the retained returns. scripts/saturation_ledger.py refuted_by_absence already treats those as not adjudicated, and landscape-sweep/build_inputs.py applies that rule. But lane_packets --manifest-newcomers withheld every refuted_* row, so 23 of them could never reach the next verdict wave (the other 4 are already ledger candidates).

The change (tools/sota-convergence/lane_packets.py):

• absence_refuted(root, manifest)
  • Loads the ledger rule and binds only completed sweeps whose manifest_sha256 equals these manifest bytes.
  • Records (lane, identity) under (catalog, layer_id) for every entry that refuted_by_absence reads as absence-only.
• manifest_newcomers no longer counts such a row as refuted.
• Unchanged:
  • a row refuted on merit still withholds the repository, and so does a mixed row (missing GPT-6 beside a returned refuting vote);
  • the default build without --manifest-newcomers;
  • the 2026-09-22 and 2026-09-23 builds.
• The rule is reused, not reimplemented. scripts/saturation_ledger.py is loaded by path, as landscape-sweep/sweep_common.ledger_module does. An import statement would pull the ledger into the verdict-review gate's trust base, which SeparationTests forbids.

Effect on the 2026-09-26 manifest: 23 more newcomers across 9 foundation packets.

SOTA sources

• Rule source, reused unchanged from this repository:
  • scripts/saturation_ledger.py refutes_on_merit, refuted_by_absence and ref_resolver (lines 236-275);
  • how it is applied: tools/sota-convergence/landscape-sweep/build_inputs.py:145-154;
  • the loading pattern: tools/sota-convergence/landscape-sweep/sweep_common.py:103-108.
• The writer that sets the missing marker: tools/sota-convergence/landscape-sweep/convert.py:452-510.
• Python stdlib importlib.util.spec_from_file_location / module_from_spec (https://docs.python.org/3/library/importlib.html#importing-a-source-file-directly). No new dependency.

Checks (at 11b5341f, rebased on main f0219263)

Check	Result
python3 scripts/validate.py	exit 0, status: passed
python3 -m unittest tests.test_lane_packets tests.test_saturation_ledger tests.test_verdict_review_gate	Ran 353, OK (skipped=3)
the three builder registry tests	Ran 3, OK
python3 scripts/saturation_ledger.py --check	ok: 4 sweep record(s), chain intact, bindings verified

Failing-first. Reproduced by an independent stack-verifier on a git archive of the base: the 8 new tests fail against base lane_packets.py (7 FAIL, 1 ERROR).

Mutation. The builder killed 6 mutants. After the Claude review, a 9th test was added: the real earendil-works/pi row is absence-only in native-clients but refuted on merit in agent-sdks, in the same lane. It kills a mutant that drops the per-layer key (frozenset().union(*absent.values())), which all 8 earlier tests missed.

Review. An evidence-reviewer round returned 1 should-fix and 3 nits, all repaired:

• the layer-scope test above;
• class docstrings changed to "not refuted on merit";
• README scope wording;
• the docstring states that the returns are trusted as saturation_ledger.py --check binds them.

The reviewer accepted loading by path for the trust base. A GPT-6 cross-family review follows before merge.

Lane

lane:foundation: tools/, tests/ for a foundation tool, and the hot-file protocol for manifests/evidence.json (docs/lanes.md:24, 96-131). No shared path changes meaning.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Cross-family review, one round. GPT-6 (gpt-6-astra, effort max, codex exec -s read-only) reviewed head 11b5341f and returned changes-needed: 1 must-fix and 1 should-fix. Both are repaired in f30dbfb7.

The reviewer also verified, beyond the two findings:

• 126 tests OK at head, and 7 failures plus 2 errors against the base builder;
• 7 further mutants caught;
• 28 expected-unchanged build configurations byte-identical to base, including the default and both historical manifests;
• the trust and separation tests;
• the ledger check, validate.py and the evidence registration.

Findings and repairs:

1. Must-fix: an unresolvable ref on a not-refuting vote still released the candidate. refuted_by_absence resolves only refuting votes. absence_refuted now also requires every role's vote ref (facts, fit) to resolve through the same ledger resolver before a release; otherwise the refutation stands. New subtests break the facts ref and then the fit ref. The facts case failed before the fix, as the reviewer reproduced.
2. Should-fix: no integration test covered the lens-vote exception. Added a lens_votes subtest, with the absence-only row as its positive control.

Checks at f30dbfb7 (rebased on main f0219263):

• validate.py: exit 0;
• tests.test_lane_packets, tests.test_saturation_ledger and tests.test_verdict_review_gate: Ran 353, OK (skipped=3);
• the three registry tests: OK;
• saturation_ledger.py --check: ok;
• two mutants, run on git archive copies, are each killed by ManifestNewcomersTests: dropping resolves(entry), and stripping lens_votes before refuted_by_absence;
• the real-data count is unchanged: 27 absence-only rows, 23 newly carried.

Residuals. None open from this round. Under the bounded-review rule there is one review and one repair, with no second review round.