Reconcile Bun shard calibration with measurement 74 and later corpus changes #1111

Measured incompatibility

Successful measurement74 attempt1 covers commit c128b655afa4e16fb509ba0b8d93a4bb7c850122 on ubuntu-latest with Bun1.4.0. Its assembled artifact records packages/cli/tests/adversarial-implementation.test.ts at 421949ms, above the documented420000ms target on its own.

The current corpus at main 97d7b44e5a43ecc0cdaa179103e90bffaf3f9ee1 also contains the later-added packages/cli/tests/repl-unfinished-entry.test.ts. It has no captured measurement, so the existing fallback charges it the heaviest applicable recorded value, also421949ms. No timing may be invented for it.

measured Bun corpus total:     3432032ms
later file's normal fallback:  421949ms
current applicable total:     3853981ms
documented floor: floor(3853981 / 420000) + 1 = 10
installed Bun matrix: 9 shards

Increasing the shard count alone cannot make an indivisible421949ms file finish below420000ms. This is a calibration incompatibility, not an invalid measurement record or missing partition coverage.

Boundary and evidence

Data-only import PR1110 keeps the exact assembled bytes and measured provenance. Its focused data/partition regressions pass10groups/52steps. Independent Codex Astra at signed 8d552b5c7e636bc4a5a5cf51b9da1f5f8145b180 passes the authorized artifact import and explicitly reports this separate calibration gap. No count change, new measurement, fixed-head calibration sequence or source repair was performed for this observation. A correctness CI pass does not establish the separate seven-minute claim.

The schema/discovery/partition rules remain in scripts/lib/test-weights.ts, scripts/lib/test-files.ts and scripts/lib/test-shards.ts; installed matrices remain19/16/9 in .github/workflows/ci.yml. AGENTS.md and specs/testing-spec.md define the measured count and five-consecutive-fixed-head calibration procedure. Follow that procedure rather than choosing a count by intuition or hand-editing milliseconds.

Follow-up

Establish why the measured workflow file exceeds the budget and obtain valid evidence for later-added tests. Preserve every substantive workflow journey and normal conservative fallback; do not exclude files or reduce coverage to satisfy the target. Make any bounded repair and calibration plan concrete before launching additional runs. Keep this issue open until the runtime/window/count claim has its required evidence.

Related investigations #1107 (lost oversized-prompt diagnostic) and #1108 (cache/outer serialization) remain unresolved. They may inform diagnosis; neither is asserted to cause or fix this timing observation. Creating this tracking issue starts no extra paid qualification, measurement or calibration.