bench-history: add 2 local parsec records (v1 pilot, v2 honest)

[bcarmeli]

Adds parsec to the live dashboard, as two distinct Experiment rows (v1, v2) — parsec has no CI job (see PR #465, parsec/benchmark-local-only, still open), so these are hand-built local-run records rather than CI-published ones, same mechanism as the existing bcarmeli local skillsbench baseline/opus-optimize records already on this branch.

What's in each record

v1 (run_id=20260908001, bench=parsec, tier=pilot, experiment=v1)

• The 2-task pilot from parsec-history/results/v1.
• reward_base 0.317 → reward_opt 0.326. Train-fit only — run_suite.sh pins train==val==test to the same 2 tasks, so this is not a generalization claim.
• best=seed: both candidates (cand_0001 mean 0.225, cand_0002 mean 0.157) scored below the seed and were rejected by the paired-SE gate. The delta shown is two separate n=30 measurements of the byte-identical seed (baseline vs. finalize pass) — seed-measurement noise, not an improvement.
• n=2 of the source experiment's 30 tasks; the other 28 are baseline-only (2 of 4 simulators were down during that sweep for most of them).

v2 (run_id=20260908002, bench=parsec, tier=v2, experiment=v2)

• The 10-task run from parsec-history/results/v2.
• reward_base 0.822 → reward_opt 0.962 (delta +0.139) — the honest equal-n framing (seed re-measured at n=9, errored trials dropped), not cap-evolve's own gate output (+0.086), which compared the n=9 candidate against a reused n=3 baseline measurement — a mixed-n comparison biased toward the newer number.
• No held-out test split exists yet for v2, so this is val-only, not a generalization claim, and not comparable to v1 (different tasks, different simulators, different scoring).
• The accepted diff is a tool-selection documentation fix, not a reasoning-strategy change (see parsec-history/artifacts/README.md).

Both records carry summary_url pointing at the corresponding parsec-history summary.md, so the Source column links straight to the full per-task table / cost breakdown / caveats instead of a bare bcarmeli. Neither record sets pr — parsec isn't CI-dispatchable (see PR #465's README changes), consistent with the other local-source records already on this branch.

Schema note

Built with the current record.py (origin/main, not the stale copy in the parsec-intake_v1 worktree I originally built these against) — both records now carry suite.n_scored, matching the new coverage-badge UI. Neither task set has any opt_infra:true row, so the reward figures themselves are unaffected by the fix; only the missing n_scored field needed correcting.

Verification

• records/*.json for both runs load and their suite rollup matches a hand-computed mean of the per-task rows (checked both).
• Fast-forwarded to the current origin/benchmark-history (671a39ce, includes bench: record transfer-eval-8fold skill-transfer pilot (run 20260902001) #459's transfer-eval-8fold record) before branching, so this isn't based on a stale snapshot.
• benchmarks.json/meta.json regenerated via record.py aggregate against the full records/ dir: 81 records, 75 runs (79 pre-existing + these 2).
• Confirmed against the live site/benchmarks.js (PR site(benchmarks): derive Benchmark/Type/Experiment filters, add Experiment column #457's Experiment column/filter): both records carry experiment so they'll show up under a v1/v2 filter value once bench=parsec is selected, no site code change needed.

🤖 Generated with Claude Code