COLEMAN 4.13 — Follow CI configurations from source data to evidence¶

Run a real, bounded LibSSH configuration history through COLEMAN's existing Whole Test Set and Variant Test Set strategies. Keep the original pipeline and job identities, inspect which configurations have usable observations, and resume experiments across both policies and budgets without losing a budget's results to another budget's checkpoint.

A complete variant-aware data path¶

The HCS guide covers the reviewed original downloads, Parquet export, library/CLI execution and offline replay. Its frozen first-sixteen pipeline slice retains 885 source rows, projects 856 test executions, and explicitly excludes 29 records from another attempt. Thirteen known job labels produce a 208-cell sparse matrix: seven labels have parsed tests, and six remain metadata-only evidence. Forty-four cells contain retained tests. Metadata-only labels, including static-analysis, do not become fabricated test configurations or executions. WTS and VTS retain their own policy state and result denominators.

The original data remain outside Git under their published CC0 dedication. Typed receipts bind source versions, exact byte counts, SHA-256, component terms, conversion code and Parquet writer. Deterministic exports publish completion last; offline validation regenerates the expected bytes from verified source. Identical concurrent retries converge, while interruptions cannot certify a partial export.

Budgets and identities you can inspect¶

Raw pipeline/job identity avoids the duplicated commit-level keys in the publisher's prepared files. Original variant labels map reversibly to safe aliases, preserving punctuation and case differences. Absent parsed cells remain absent evidence rather than fabricated passing tests.

The new recipe explicitly aligns modeled costs with the existing bandit's binary32 duration precision before deriving budgets. Original decimal seconds and their lexemes remain unchanged in analytical Parquet. Union and per-variant costs therefore have one declared interpretation; this is test-only serial simulation, not measured runner occupancy or physical savings.

Native multi-budget checkpoint namespaces now include the exact mode/value. Single-budget paths and payloads remain compatible. Ambiguous old unscoped multi-budget state rejects before loading, including nested extension run IDs and legacy names resembling the new namespace. Bounded progress metadata distinguishes scoped state without unpickling checkpoints. No migration guesses which budget produced old results. Existing native data, source recipes and artifacts retain their meanings.

Native local Parquet recovery also closes the gap between flushing results and committing a checkpoint: resume reconciles the uncommitted tail before replay, including an interruption before the first checkpoint. Complete shards publish atomically; failed writes cannot expose partial Parquet as completed results. Committed rows and other executions/budgets remain intact. Same-run attempts must be serialized; this process-interruption guarantee does not imply power-loss durability or equivalent behavior for every optional sink.

Persistent ownership bindings also prevent a fresh or emptied checkpoint directory from erasing another run's results at matching dataset/budget keys. Ambiguous old tails reject; completed legacy coordinates gain ownership metadata after their loaded checkpoint is verified, preserving their original rows and payloads. Custom providers keep their existing checkpoint behavior.

Evidence before claims¶

The slice exercises Random/UCB with RNFail, ratio budgets 0.5/1.0 and checkpoint continuation. Native zero reward means nondetection, not an observed passing label. Upstream binary nonpass labels may include statuses already collapsed by extraction; they are not independently mapped unique faults.

Variants, repeated pipelines and seeds do not create independent projects or faults. Outcome/feature availability, upstream omissions, setup, shared occupancy and concurrency remain explicit evidence gaps. This release advances public

261 source reproduction; scientific admission, protected comparisons and the¶

confirmatory cross-corpus report remain open.