fix(bench): modern DataHub seeding, enrichment-volume reporting, and the first published pilot results

[cjimti]

Part of #930. Fixes two bench-harness defects surfaced by standing the a2 arm up on a current DataHub, and starts the living results page that the benchmark phases (#943-#945) will grow, seeded with the phase 1 pilot numbers.

DataHub seeding works on current CLIs

The MCE emitter wrote the legacy datahub put --file bulk format, which no longer exists in current datahub CLIs (put is now a subcommand group requiring --urn). The emitter now writes the MCP file format the datahub ingest file source consumes (each aspect wrapped as a GenericAspect {"json": {...}}), the seed artifacts are regenerated, and make bench-seed-datahub generates a file-source recipe and drives datahub ingest against BENCH_DATAHUB_GMS (default http://localhost:8080). Verified live: 12 events ingested into a DataHub v1.6.0 quickstart running on OpenSearch.

Enrichment volume is no longer invisible in reports

Suite reports summed enrichment_tokens_dedup, which is zero when enrichment runs in full mode, so the a2 arm's enrichment volume displayed as zero even while audit rows showed enrichment_applied: true with per-call token counts. auditapi.Metrics and the suite summaries now carry enrichment_tokens_full (the delivered volume) alongside the dedup counter, and the human summary prints the full volume. Verified live: the a2 pilot run reports ~7.3k enrichment tokens across the knowledge suite.

Published results: a living page plus committed history

docs/reference/benchmarks.md is the new reference page (mkdocs nav and the docs/llms.txt / docs/llms-full.txt pair updated). It states the reading rule (arm-vs-arm on a pinned model; the model is disclosed, never the subject), and publishes the phase 1 pilot: manifest, arm-by-suite tables, per-trap-class breakdown, honest caveats, and a result-history table each future phase appends to. The raw run JSON (claude-sonnet-5, k=3, 60 episodes, zero harness failures) is committed under bench/results/v1.102.0-pilot/ as the first history entry, per the #930 results-history convention.

Pilot headline, also recorded on #930: knowledge-trap accuracy 60% (a0, raw tools) vs 100% (a2, platform), with the two tasks whose disambiguating fact lives only in the knowledge layer going 0/6 vs 6/6 while the baseline exhausted its 30-call budget on every attempt; knowledge-suite median tool calls 16 vs 8; median wall clock 67s vs 25s. Caveats are published alongside: both arms score 100% on trivial discovery (where a2 pays a small search-first friction cost), and two trap tasks fell to raw model inference; the page marks all pilot numbers as directional pending phase 2's larger suites and judge calibration.

Verification

• make verify green on this branch (root gates, including doc-sync checks over the new docs).
• Bench module gates green: build, vet, tests, golangci-lint.
• Both fixes exercised against a live stack: seeding into DataHub v1.6.0/OpenSearch, then a full paid a2 run (30 episodes) whose report shows the enrichment volume; the a0/a2 result pair committed here is that run.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.04%. Comparing base (1984dd2) to head (ac20012).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #951      +/-   ##
==========================================
+ Coverage   89.03%   89.04%   +0.01%     
==========================================
  Files         424      424              
  Lines       48581    48581              
==========================================
+ Hits        43253    43260       +7     
+ Misses       3551     3547       -4     
+ Partials     1777     1774       -3     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.