CI witness cost backout: trim the floor's enrolled roster by measured cost (103 -> 67 check fns)

[briansrls]

CI witness cost backout: trim the floor roster by measured cost, 103 → 67 check fns

Operator ruling 2026-07-10: "CI taking 1 hour is way too long … we'll need to back most out until we get performance under control."

How witnesses get into CI (the answer to "how are those selected")

Single authority: commit_gate_roster in dag/gunbc/commit_workflow.dag. Each CommitWitnessClaim { entry, check_fns } row on the GithubActionsCiJob surface is projected into the floor plan by project_ci_floor_witness_entries at floor runtime (receipt: this PR's regen produced zero ci.yml drift). Enrollment is opt-in (operator inversion 2026-07-04, which zeroed it); it accreted 0 → 103 check fns in six days as each merged lane enrolled its own witnesses.

Where the hour went (measured, run 29108981678 — the 1h0m50s ci job)

segment	time
release-bin build (cache+sccache warm; PR carried main's new .rs)	26m52s
batch 1 — compile-clean gate	2m22s
batch 2 — 103 witnesses + drift/layering gates	15m06s
batch 3 — emit_host / ingest gates	2m01s
batch 4 — emit_determinism gate (one gate)	13m31s

Witness count ≠ witness cost: p50 witness = 294ms. Of batch 2's 15m, ONE witness (dag_collect_fingerprint_witness_keystone_holds, #6425) was 8m38s; five heavies ~2.7m; the complexity-lens ~12s-eval family ~3.5m; the ~80-witness tail costs under a minute combined.

The cut (by cost, not count)

• Unenrolled dag_collect_fingerprint_witness_test (518s keystone).
• Dropped parse_table_memo_warm_equals_cold_content_hash (36s); the row's four governance RED controls stay.
• Trimmed my five complexity/body-lowering suites from 39 fns to RED control + flagship positive each (10 fns) — accumulator_copy_fold_analysis, complexity_accumulator_copy, fold_lowering, loop_multiplicity_measure, loop_infer_iteration.
• Deliberately KEPT (~101s total): enforcement_live_witness_test (84s) and budget_roster_completeness (17s) — the §6 "lenses must be live" backstops that keep every other lens honest. Highest displaced-cost-per-second on the roster; overrule if you want them out too.
• Left the sub-second tail enrolled (materialization suites etc., ~44 fns): cutting them saves <1m combined and they're the cheapest RED walls we have.

Every trimmed fn stays in its witness file and still runs on the local discovery path (claim_batch --roster-from-discovery …); existence ≠ enrollment. Re-enrollment is a one-row edit when parse/resolve memoization lands.

Expected effect: batch 2 ≈ 15m → ~2.5m.

What witness backout cannot fix (flagged for their own lanes)

1. Release-bin build: 26m52s on any PR carrying .rs changes (sccache-warm — release codegen/link-bound). Lever: reuse a prebuilt binary artifact keyed on .rs content-hash (the same Realization/materialization argument as everything else). No .rs delta → 1m44s (measured on main).
2. Batch-4 emit_determinism gate: 13m31s, a single structural gate, not a witness entry.
3. The floor shows ~2× variance under runner contention (same corpus: 17m on main vs 33m on the concurrent PR run).

With this PR, a typical no-.rs-change PR's ci job ≈ 2m build + ~9m floor. With .rs changes it's still ~27m build + ~9m floor until lever 1 lands.

Durable fix candidate (follow-up, not this PR): enrollment rows declare a per-fn eval budget and the floor REDs a witness that exceeds it — cost-gated enrollment as a wall, so the roster can't silently re-accrete a 518s witness.

Receipts

• main_wet regen: zero generated-artifact drift; drift gate PASS.
• Kept-witness sanity run: quadratic_fold_has_suspect PASS wet against the trimmed roster entry.
• Diff vs main: exactly one file, dag/gunbc/commit_workflow.dag (+5/−44).

🤖 Generated with Claude Code

[gunbai-bot]

The "CI FAILING (3 failing)" dashboard reading is the known superseded-run artifact, not a real failure: the draft-to-ready flip re-triggered CI under the concurrency group, cancelling the first run on this same head (29113213972, cancelled) and the WIP-head run (29113098355, cancelled). The live run is 29113261146 — ci_regen already green, ci and rust_tests in progress. No fix needed; this run doubles as the measured receipt for the trimmed floor. — sent from eager-ram-612