Repository navigation
CI budgets: repair #6246 rebase regression (rust gate 10m->45m, warm 15m->45m, derived backstops) - #6338
Conversation
…ust gate/warm values, re-derive backstops The #6246 rebase restored OLD budget values while keeping the NEW notes that document larger receipt-backed values: - gunbc_ci_rust_gate_step_timeout_minutes: 10 -> 45. Its own note: nextest killed mid-run at the 30m cap (receipt job 85319650284: warm 14m38s, reconcile 1339s before main, nextest in-flight at kill); prior 30m receipt: clippy in-flight at 15m kill. - gunbc_ci_rust_sccache_warm_step_timeout_minutes: 15 -> 45. Its own note: PRs 6286/6274/6290 timed out at 15m under cold-cache fallback. Live consequence repaired: every rust_tests run false-killed at the 10m gate step (receipt: PR #6337 job 85487003660, step timeout at exactly 10m, zero test output). Values were one config edit below their own receipts. ci.yml + falsifier.yml regenerated via generated_artifact_gate::main_wet - derived values move together (rust_tests job backstop 40 -> 105, falsifier job 155 -> 170 + its stale build-command drift repaired by the same regen). Deliberately NOT swept in from the same regen (pre-existing drift, reported separately): DESIGN.md - regenerating it DROPS the operator's 2026-07-05 section-5 ruling text (escape hatches / factory model / review bar) because design_document.dag is stale behind the hand-authored authority doc; and a .gitignore exception line. Both left at their committed state. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
ee817f4 to
1ed2bba
Compare
…derive disposition prose to 106m sum Both cursor findings verified real and fixed: - warm 45->46: witness_warm_step_budgeted_separately_from_gate asserts warm strictly > gate; witness_rust_tests_yaml_has_observable_warm_step_before_gate pins the emitted 'timeout-minutes: 46' before the gate step. Both now green by execution. git -S receipt: #6246's own branch carried 'fix(ci): restore rust gate step budgets that regressed to 10m' (303d379, warm=46/gate=45) - the squash dropped it; this restores the intended state. - disposition prose 40m/15+10+5+5+5 -> 106m/46+45+5+5+5 (single authority: prose re-derived with the carriers, warm>gate separation named). ci.yml regenerated via main_wet (warm 46, derived rust_tests backstop 106). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Composer review (36107) — both findings verified and fixed in
|
…ved, keep 6339 gate-note drift history Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Post-#6339 delta (after merging main): #6339 restored gate=45 only and reproduced the warm prose-data drift (warm left at 15 while its own note says 45m) — main is still red on the enrolled serializer witness pinning emitted timeout-minutes: 46 before the gate step. This PR's remaining delta: warm=46 (witness-pinned, strictly > gate), backstop re-derived 106m = 46+45+5+5+5, falsifier.yml drift repair (45m build, 170m job, --bin narrowing). #6339's gate-note drift-history addendum kept verbatim. Serializer witnesses green by execution on the merged tree. Note for the record: an auto-commit pushed the mid-resolution tree with conflict markers as 7de0be5; resolved and superseded by the follow-up commit — do not review 7de0be5 in isolation. — sent from sunny-wren-799 |
|
CI status disposition: the failing check is the — sent from sunny-wren-799 |
What
The #6246 rebase restored old CI budget values while keeping the new notes that document larger, receipt-backed values — an authority self-contradiction (value vs its own note):
gunbc_ci_rust_gate_step_timeout_minutesgunbc_ci_rust_sccache_warm_step_timeout_minutesLive consequence: every
rust_testsrun now false-kills at the 10m gate step. Receipt: PR #6337 job 85487003660 — step timeout at exactly 10m00s, zero test output survived the kill. Main has no post-#6246 run yet; its next one hits the same wall.Fix
Values restored to their own notes' numbers (45/45);
ci.yml+falsifier.ymlregenerated viagenerated_artifact_gate::main_wetso derived values move together: rust_tests job backstop 40→105, falsifier job 155→170 (+ falsifier's stale build-command drift repaired by the same regen). Floor-step 10m and the era-5 step-level semantics from #6246 are left untouched — this repairs only the two values that contradict their own receipts.Deliberately excluded (found by the same regen, reported to operator separately)
design_document.dagis stale behind the hand-authored authority doc. Shipping that regen would regress an operator ruling; left at committed state pending an operator re-sync ruling..gitignoreexception line (unrelated drift).🤖 Generated with Claude Code