feat(benchmark): add the autonomousqa member [roadmap:qa-benchmark]

[tcballard]

Implements the qa-benchmark roadmap (lore-proofkeeper/roadmaps/qa-benchmark.md in itsthelore/proofkeeper): a public, deterministic, agent-agnostic benchmark for autonomous-QA agents, landing as the autonomousqa/ member alongside decisiongrounding/ per ADR-092.

What's in the member

• Four frozen sample apps, one per drive modality, all Python-stdlib served with zero dependencies to drift: browser-notes (browser flow), api-ledger (API service), cli-tally (CLI tool), ext-wordbadge (MV3 browser extension). Each seeds a Lore corpus that validates clean, grading capabilities easy → hard, including negative paths and deliberately ambiguous capabilities whose honest outcome is unverified.
• Harness — takes (app, corpus, agent config); exports the corpus over the published rac export --graph contract; invokes the reference agent over the published @itsthelore/proofkeeper npm CLI (pinned 2026.7.1 in workspace/), never repo internals; and writes append-only records with verified/unverified per capability, fidelity pass-rates, tokens in/out, and wall-clock. The agent invocation is a pluggable seam (agents.base.AgentAdapter) so a second agent is one adapter away.
• Token metering and the CI stub are one component: an OpenAI-compatible proxy on the agent's documented BYOK seam (OPENAI_BASE_URL). Forward mode meters real provider usage per call; scripted mode replays canned per-capability flows so CI runs the whole pipeline — drive, compile, fidelity gate, scoring — with no key and no token spend.
• Deterministic scoring (ADR-066 lineage): verdicts re-derive from recorded raw evidence (stdout + exit code) with a pure parser — no embeddings, no LLM judge; recorded results re-score offline without re-spending tokens. Exit 0 without fidelity evidence scores as an error, never a verification. Honesty is scored: verifying a capability seeded ambiguous counts as a false verification.
• Results page generated from records: rates by app, by modality, by model, by tier; the exact harness config on every row; run-to-run variance for repeated configs. The shipped result set is a scripted-run illustration, labelled as such.
• CI (.github/workflows/autonomousqa.yml, at repo root with path filters since workflows don't execute from member subdirs): offline pytest matrix (3.11–3.13), rac corpus gates over the member corpus and every seeded app corpus, and the smoke job — one cheap capability against one app with the scripted model, through to a deterministic double re-score.
• Member corpus (autonomousqa/rac/): the requirement, the harness design, and two ADRs (metering/stub over the BYOK seam; strict evidence-based scoring), gated in CI.

Evidence

• All 19 capabilities seeded verifiable verify end-to-end through the real pipeline with the scripted model, across all four modalities (make fixtures).
• 47 offline tests; rac gate rac passes; the workflow is green on this branch (run 28569902233).

Notes for review

• The published CLI's npm bin shim was observed to silently no-op (exit 0) when invoked via symlink; the adapter therefore invokes dist/cli.js by real path, and scoring refuses to count exit 0 without fidelity evidence — both recorded in ADR-0002.
• Apps and corpora freeze on publication; behaviour changes land as new apps (freeze policy in the member README).