📊 Backfill A/B measurement on shipped doctrine variants (depends on #131)

[ZaxShen]

Once #131 ships the A/B framework, use it to retroactively measure already-shipped doctrine variants. Stop guessing about the LLM compliance ceiling; get data.

Hypotheses to test (each ≥10 paired runs per arm)

1. CLAUDE.md slim: did 268 → 109 lines actually improve outcome pass-rate?
   • Arms: pre-slim (snapshot from before PR ✂️ refactor(claude.md): aggressive slim 142→99 lines, extract every protocol to skills/docs #126) vs current
   • Metric: outcome pass-rate per flow + total tokens per session
2. Hybrid D' (cold-start AskUserQuestion + lazy default): better than pure-lazy or pure-eager?
   • 3 arms: Hybrid D', always-lazy, always-eager
   • Metric: outcome pass-rate + tokens + time-to-first-useful-action
3. Direct Mode 4-step protocol: did the strong NEVER-SKIP-THIS framing move compliance vs. prior 3-step + soft framing?
   • Arms: pre-PR 🐛 fix(bro): tighten activation routine + Direct Mode protocol; fix Task→Agent scorer mismatch #139 wording, current 4-step, hypothetical post-commit-hook enforcement
   • Metric: rate at which direct_mode_used ledger event lands
4. First-action chain MANDATORY tightening: did PR 🐛 fix(bro): tighten activation routine + Direct Mode protocol; fix Task→Agent scorer mismatch #139 move greeting-flow compliance, or are we at the LLM ceiling?
   • Arms: pre-PR 🐛 fix(bro): tighten activation routine + Direct Mode protocol; fix Task→Agent scorer mismatch #139, current
   • Metric: rate identity_get + issue_resume appear in trajectory on @bro hi

Why this matters

Several recent doctrine PRs were guesses. Without comparison runs we don't know which helped. If hypothesis 4 confirms the ceiling is real, we stop prompt-only fixes for compliance and consider programmatic enforcement.

Acceptance

• Each hypothesis run with ≥10 paired runs per arm
• Results published as ADR under docs/trustmybot/architecture/manual/decisions/
• If a variant lost or was within noise, file followup to revert or re-design
• Findings inform v0.4.3+ prompt shipping

Sequencing

Blocked by #131.