experiment-tracker: reps carry no evidence pointers (run id / artifact refs), so lineage cannot reach the record

[drewstone]

Consumer story

discovery-lab keeps a register of scientific results (RESULTS.md) whose every entry must resolve mechanically to runs, oracle grades, and artifacts. The lab built a provenance graph (tools/provenance.mjs) where a register result cites runs and every answer carries source file paths. It wanted to adopt experiment-tracker.ts as the experiment/lineage substrate and could not: the chain from an experiment to the evidence that produced its scores does not exist in the type.

What was read (0.145.3 published package; src/experiment-tracker.ts)

• ExperimentRep is {rep, score, timestamp, passed?, metrics?} — a bare number with no run id, no result path, no artifact pointer. The only file paths in the module are ExperimentProvenance.changedFiles (git-diff paths).
• Experiment.parentId gives experiment-to-experiment lineage, but a rep cannot name the run or record it measures, so "which run produced this score" is unanswerable from the tracker alone.
• Contrast: the package already has the right vocabulary elsewhere — SearchArtifactRef {role, uri, sha256, byteLength} (campaign/search-ledger.ts) and EvidenceRef {kind, uri, excerpt?} (analyst/types.ts). Neither is reachable from ExperimentRep.
• fileExperimentStore is a whole-file JSON rewrite with no append, chain, or lock, while ledger-core owns exactly those properties.

Required change

ExperimentRep accepts optional evidence: a runId?: string and evidence?: EvidenceRef[] (or SearchArtifactRef[]) that validation, stats, and verdicts preserve unchanged. computeExperimentStats/improvementVerdict ignore them; stores round-trip them.

What this deletes downstream

discovery-lab's register→run join stays in RESULTS.md prose + a local registry file today; with rep-level evidence refs, register results become Experiments whose reps point at runs, and the lab's registerChain query becomes a projection of upstream records. Recorded in discovery docs/15-adoption-decisions.md.

Note on adopted surfaces (working well)

The same lab work adopted supervisor-run (readRuntimeSupervisorRun, analyzeSupervisorRunSources, journal dialect detection) and ledger-core (hashCanonical) directly over real run directories with zero adapters — those fit exactly.