feat: write experiments.jsonl per run with action workflow run_id

[Copilot]

Experiment state had no durable, run-correlated JSONL log. assignments.json has no run_id; state.json embeds a runs[] array but is primarily a counter store. Neither is convenient for correlating experiment assignments with a specific GitHub Actions run.

New file: experiments.jsonl

A growing JSONL file committed to the experiments git branch alongside state.json. Each line:

{"run_id":"12345678","timestamp":"2026-07-15T18:00:00.000Z","assignments":{"style":"detailed"}}

Pruned to the last 512 entries (matching MAX_RUN_HISTORY).

Changes

• pick_experiment.cjs — after picking variants, appends a {run_id, timestamp, assignments} record to experiments.jsonl in the state dir; prunes to 512. runRecord is constructed once and shared with state.runs to avoid duplication.
• load_experiment_state_from_repo.cjs — fetches experiments.jsonl from the experiments git branch alongside state.json so pick_experiment can append to the existing log (non-fatal on 404).
• compiler_experiments.go — adds GH_AW_STATE_FILES: state.json,assignments.json,experiments.jsonl to the push_experiments_state job so all three files are committed to the branch.
• pkg/constants/job_constants.go — adds ExperimentsRunsFilename = "experiments.jsonl" constant (mirrors EvalsResultFilename).
• experiments-specification.md — adds experiments.jsonl to the definitions table; updates R-STORE-REPO-001/003; adds R-STORE-REPO-006 specifying write/prune behaviour.
• pick_experiment.test.cjs — 4 new tests: initial write with correct run_id, no write when no assignments, append across successive runs, pruning at MAX_RUN_HISTORY.
• *.lock.yml — all 256 workflows recompiled.

[github-actions]

🤖 PR Triage

Field	Value
Category	feature
Risk	🟡 Medium
Priority Score	72/100
Action	⚡ fast_track

Score breakdown: Impact 35/50 · Urgency 20/30 · Quality 17/20

Rationale: Adds durable experiments.jsonl logging alongside state files. Wide blast radius (49 files, 256 workflow lock files regenerated). Well-described with 4 new tests. CI not yet run. Medium risk.

Generated by 🔧 PR Triage Agent · 36.7 AIC · ⌖ 7.39 AIC · ⊞ 5.6K · ◷