Seed the witness row-cost basis so the drift comparator can fire

[gunbai-bot]

dag/gunbc/witness_row_cost_basis.tsv has carried a header, two comments, and zero data rows since it was created. So every floor run printed:

[receipt] floor witness row-cost drift: basis_absent=5344 drift_exceeded=0

That reads as "no drift". It is the comparator being structurally incapable of reporting any — drift_exceeded was 0 by construction, not by measurement.

The fail-quiet is the sharp part. claim_executor has a loud witness-row-cost basis file missing at {} diagnostic, but it only fires when the read fails. The file existed and was merely empty of data, so the loud arm was skipped and nothing counted 5,344 BasisAbsent rows as a deficit. This is the mechanism that would have warned as resolution_divergence_silent_pick_gate_keystone_holds climbed 717s → 815s → 900s into its ceiling. It went dark instead, and the row crossed with no advance signal.

Where the rows come from

Falsifier run 30550328673 — runner_spec_is_self_hosted_arm64, the host class witness_row_cost_basis_host_class_note requires ("never local x86"). The rows are read from that run's own [witness-row-cost] receipt lines, which claim_executor eprintlns per row. Its sole failure was the silent-pick row's wall-budget refusal; every seeded row completed normally.

• 1,192 rows seeded
• 4,148 zero-eval rows correctly unseedable (the parser already refuses eval_ms_basis == 0)
• 3 keys observed in two batches keep the max of the pair (all differ by <2% — themselves an instance of the duplicate-work thread, the same witness evaluated in batches 6 and 7)

The row deliberately left out

resolution_divergence_silent_pick_gate_keystone_holds — the corpus's most expensive witness — is not seeded, and that is the load-bearing decision in this PR.

It was killed at its deadline. The log carries wall budget exceeded: 900001ms elapsed > 900000ms whole-receipt budget, yet its row-cost entry records a flat 900794 ms, indistinguishable from a witness that genuinely finished in that time. That is a censored measurement rendered as a completed one — DESIGN §5's ⊤-as-answer conflated with ⊤-as-ignorance.

Seeding it would be worse than useless. Under the 2× comparator (witness_row_cost_exceeds_basis), a basis of 900794 means drift cannot fire below ~1.8M ms — so the row most in need of a cost signal would be pinned WithinBasis forever, and a deadline ceiling would become its official normal cost. That is the re-basis trap, arriving through the basis file instead of through the budget constant.

It stays BasisAbsent, which is the honest third state, until ClaimOutcome::TimedOut can distinguish a killed row from a completed one. This makes the TimedOut plumbing a prerequisite for a correct basis on exactly the expensive rows, not independent cleanup.

Cannot red CI

The drift receipt writer returns true regardless of drift_count and fails only on a write error, so drift stays observational — matching witness_row_cost_spine_note's "falsifier-cadence counted, not merge-refusing".

Verified by execution

committed_witness_row_cost_basis_loads_and_excludes_the_censored_row runs the real committed artifact through the loader's own parse_witness_row_cost_basis_line. It asserts ≥1000 rows load and panics on any row the loader would refuse — the loader silently skips refused rows, so a malformed commit would otherwise degrade back to BasisAbsent unnoticed, which is the exact failure mode this PR exists to end.

Discriminating RED controls, both directions:

control	result
basis restored to the old vacuous state (0 data rows)	FAILED — committed basis loaded only 0 row(s)
censored row re-seeded at its 900794ms ceiling	FAILED — names the offending row
the real committed basis	ok

Full claim_executor suite: 36 passed, 0 failed.

Related

Discharges the reproduction blocker named in placement_workstream_task_20_open_unknown_note (status OpenUnknown), whose instructions require per-witness cost receipts from the arm64 fleet host class.