CI floor step window 60→90: post-converge composition no longer fits (two labeled receipts)

[gunbai-bot]

At the converged 16 GiB slot caps (memory.max read in-run on both srv1 and srv2), the ci floor step no longer fits its 60-minute window even under healthy selection — two labeled receipts from tonight:

• Run 29140422757 (PR regen_stage0: runtime workspace_root() instead of baked CARGO_MANIFEST_DIR (defer to #6484 authority) #6482, pure-.rs diff): batch-1 compile-clean + a 26-minute selection walk that skipped 1,535 rows (selection working; skip-before-resolve means every skipped row still pays its resolve) + the .rs gates (nextest) unfinished at +27 min → step killed at 60, floor_peak_post=16106758144 censored (srv2-02).
• PR Host-scaffold classifier defork: substrate-declared LiveTreeDisposition; delete the Rust text classifier #6479's run: batch-1 ~21.5 min, corpus batch still executing at the kill, floor_peak_post=16140107776 censored (srv1).

Time was the censor in both; memory is the true wall (~1 GiB of margin with rows remaining). At 90 a run reaches either completion (exact-labeled) or the cap kill (censored-labeled) — both are measurements, per the same discipline as the falsifier raise (#6485). Single authority gunbc_ci_floor_step_timeout_minutes 60→90 with a receipt note; job backstops derive so they stretch in step; ci.yml regenerated (both floor and emit_determinism steps carry the value). The note names the dissolve-on: revisit down when S2a move 2 (module split) shrinks cold-resolve wall+residency and skip-before-resolve lands.

Companion to #6482 (the fleet-unblock fix this run's receipt came from) and #6485 (falsifier-side raise).

🤖 Generated with Claude Code

[gunbai-bot]

ci red @ 5a1788d is the inherited cross-slot baked-path class (job on srv2-05 drew a binary carrying srv2-03's baked workspace root — cli_run.rs:379 panic at floor startup), not this diff. Fixed by #6482; rerun after it merges. Same classification as #6485's red. — sent from loyal-wren-398

[gunbai-bot]

Merge-order note for the operator: this window raise (60→90) is what unmasks the interp_recorded_fixture_witness dark red (6/23 checks, red on clean main since #6210, currently hidden by the batch-2 timeout blindness). The fix rides #6485 (out-of-pool entry refusal + witness scratch re-home, ea8a26b). Merge #6485 before (or together with) this one, or main's first 90-min run will show 6 witness FAILs that are pre-existing, not caused by this diff.

— sent from loyal-wren-398

[gunbai-bot]

Operator synthesis — this PR is the keystone for a main-red condition, cross-referencing four blocked PRs.

Main is RED now on the 60-min floor-step timeout: run 29157805332 @ 3a3da70 (pure post-#6484 main, none of my diffs) ci=failure. Broad-footprint floors (anything touching cli_run.rs / 04_infer / 00_core → ~whole-corpus affected set) exceed 60 min.

Three-part main-green fix, all at/near the merge bar:

• Enforcement coverage witnesses offline: exclude the 3 live-tree corpus-scan rows from per-PR CI (operator ruling 2026-07-11) #6488 corpus reduction (offline 3 live-tree scan rows) — MERGED; its own ci run 29160021454 is queued (fleet-saturated), so the reduced-corpus floor time is not yet measured.
• CI floor step window 60→90: post-converge composition no longer fits (two labeled receipts) #6486 (this PR) floor step 60→90, interim window — 2/2 approvals.
• Falsifier cadence follow-ups: step timeout 120→170 (measurement was time-censored) + un-orphan 13 plan docs #6485 interp_recorded dark red (6 checks) that becomes the visible terminal failure once the timeout lifts — 2/2 approvals.

Blocked-by-systemic, not by defect — all at 2/2 approvals, all timing out on this same floor cost: #6482 (~47 min into its 60-min window as I write, likely to time out), #6485, and jolly-swift's #6480 / #6487 (byte-identical + 773-green local receipts). None is a diff defect.

Recommendation: merge #6486 first (extends the window), then the broad PRs merge current main — picking up #6488 + the 90-min ci.yml — and rerun once. If a whole-corpus floor still exceeds 90 after #6488's reduction, that's the merge-past-with-local-receipts case (the #6438/#6453 precedent); I'll bring #6488's run-timing the moment it concludes so the width>1 vs window-vs-merge-past call has data. Note: neither the skip-before-resolve nor the interface-grain lane shrinks a whole-corpus floor — nothing to skip when everything is affected — so those don't self-unblock these broad PRs.

— sent from loyal-wren-398

[gunbai-bot]

Closing as SUPERSEDED — and correcting my own earlier framing of this PR as the CI keystone.

Why closed: this PR raises the floor step timeout 60→90. But main has since escalated it to 270 minutes (gunbc_ci_floor_step_timeout_minutes: Duration = 270 in both dag/gunbc/ci_workflow.dag and the generated ci.yml, no drift) through a chain of raises chasing the full-corpus floor cost: #6459→90 (body_lowering normalize hook), then →120→180→270 (#6464, after the 180m cap timed out mid-batch-2). Merging this PR would REGRESS 270→90 and kill every floor run. Its only content main doesn't already have is the converge_raise_note, whose 60→90 "converge to 16GiB / memory is the true wall" story is subsumed by main's discovery_flip_note (which records the →90 step and every raise past it).

The larger point: the window-escalation approach is exhausted. main is at 270 min (4.5h) and the escalation ledger shows even 180 timed out mid-batch-2 — the floor's full-corpus witness-execution phase genuinely runs for hours, and chasing it with a bigger timeout-minutes is the unbounded-cost treadmill DESIGN §1/§6 warns against, not a fix. The real levers are:

• corpus-reduction — Enforcement coverage witnesses offline: exclude the 3 live-tree corpus-scan rows from per-PR CI (operator ruling 2026-07-11) #6488 (offline 3 live-tree scan rows), MERGED;
• affected-set skip-before-resolve — W2 (skipped rows still pay resolve today; the ~26-min walk over 1,535 skips is pure waste);
• width>1 parallelism — the floor-shared-compute-memoization lane;
• merge-past with local receipts — for individual broad-footprint PRs at the merge bar (the Turn on affected-set CI: re-land witness enrollment flip (discovery shrunk by affected set) + falsifier host-OOM receipt #6438/affected set processing #6453 precedent).

A future timeout lowering (once #6488's reduction proves the floor fits under 270) would be a fresh, receipt-backed PR — not this stale one.

— sent from loyal-wren-398

[gunbai-bot]

Re the conflict-marker REQUEST_CHANGES: correct finding, but this PR is now CLOSED as superseded (see the close comment). The markers came from an auto-committer concluding a mid-merge state; I did resolve them locally, but the PR itself is obsolete — main's floor timeout has escalated to 270 min, so this 60→90 PR would REGRESS it. Nothing to fix here; the marker branch won't merge. — sent from loyal-wren-398