feat(benchmark): leaderboard artifacts as spec-conformant EvalScoreRecords (#9265)

[JSONbored]

Stacked on #9600 (stack #9601): #9595 → #9598 → #9599 → #9600 → this.

What

#9216 requirement 5 is explicit that leaderboard artifacts are exactly the eval-interface's consumable score records — no parallel format. So this emits EvalScoreRecords with workUnit.kind === "benchmark_run" through the same contentDigest path the outcome_confirmed_precision records already use, and derives the board from those records and nothing else.

That last point is the structural one: there is no second query path from the scorer to the board. A number on the leaderboard is, by construction, a field of a record a third party can fetch, re-derive and digest-verify — so the published board and the consumable feed cannot disagree, because they're the same data read twice.

Two policies enforced here rather than left to a caller's discipline

Held-out cadence (#9263 decision 2). The emitter is the last point before a number becomes public, so it's where cadence is enforced. A held-out record is simply not built while the window is open — not built-and-filtered, not built-with-a-flag — so there is no in-memory held-out record for a downstream bug to serialize. The regression test probes this with the would-be record's own digest, a 64-hex value unique to that exact content, because my first attempt used the held-out precision as the probe and it was 0 — a substring of almost any JSON, so it passed for the wrong reason.

Trust tier from the actual execution. attested requires a real envelope; absent one the tier is reproducible, never optimistically upgraded. The record type's discriminated union already makes "attested without an envelope" unrepresentable, and the tier sits inside the digest preimage, so a tier claim can't be edited after issuance (asserted: the two tiers produce different digests for otherwise-identical content).

Score mapping and ranking

The score block maps the multi-class report onto the spec's fixed fields without inventing any: precision/recall are #9262's macro headline, decided/abstained/coverage come from the report's own coverage block, and confirmed is the count of decided units the agent got right. Ranking is macro precision descending with null last (unmeasurable is not zero, and sorting it as 0 would invent a claim), ties broken by coverage descending then by the earlier issuedAt. Immutability is respected rather than worked around: several records for one subject resolve to the latest issuedAt, because a corrected score is a new record, never a mutation.

Tests (14, 100% of branches)

Spec conformance including an exact top-level key set (no extra fields); full commitments block with no placeholders; the score mapping; the decided-nothing null arms; both trust tiers with distinct digests; the not-built-while-open regression; publication at close and on a cadence boundary but not a day either side; the ranking comparator across every arm (null-vs-null falling through to coverage, the ?? -1 fallback from both argument positions, the earlier-claim tiebreak, order-independence); the corrected-score regression; cross-benchmark and cross-kind exclusion; and the empty set.

Also registers the module in check-dead-source-files' staged-ahead allowlist — the emitter lands before its serving route (#9216's endpoint sub-issue) and before #9264 supplies the envelopes its attested tier consumes, so the checker's flag was correct and the entry names both, as the checker itself instructs.

Closes #9265

[loopover-orb]

Important

🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨🟨

⏳ LoopOver is waiting…

LoopOver has seen this pull request and is waiting on CI checks to finish before reviewing it. This comment will update once the review runs.

_{🟩 Safe / merged · 🟦 Advisory · 🟨 Held for review · 🟥 Blocked / closed · 🟨 Waiting}

[superagent-security]

Superagent didn't find any vulnerabilities or security issues in this PR.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 90.10%. Comparing base (9942512) to head (35bf2be).
⚠️ Report is 3 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #9602   +/-   ##
=======================================
  Coverage   90.10%   90.10%           
=======================================
  Files         890      891    +1     
  Lines      112223   112251   +28     
  Branches    26610    26621   +11     
=======================================
+ Hits       101120   101148   +28     
  Misses       9773     9773           
  Partials     1330     1330           

Flag	Coverage Δ
backend	95.52% <100.00%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
src/review/benchmark-eval-records.ts	100.00% <100.00%> (ø)

[loopover-orb]

LoopOver approves — the gate is satisfied and CI is green.

LoopOver retracted this approval — a newer commit no longer qualifies.