feat(cli): DX, evidence & hardening — uniform --json, broader doctor, eval harness (Tranche D)

[CodeWithJuber]

What & why

Tranche D — developer-experience, evidence, and hardening from the review.

Uniform --json

doctor, route, preflight, verify, and scope now accept --json (only
impact/substrate/anchor did before). CI and scripts can now gate on the health check, the
routed tier, the assumption gap, and the verification result — not just parse human text.

forge doctor catches more silent misconfiguration

New checks, each targeting a real "it silently stopped working" failure mode:

• guard scripts present and executable (a non-+x guard makes a hook silently no-op),
• jq / git availability (several guards degrade without jq — this surfaces the risk the
  audit flagged for secret-redact/protect-paths),
• atlas presence + freshness (a stale graph misleads impact/verify),
• model-pricing staleness (warns when the verified date is >90 days old).

Evaluation harness (src/eval.js)

The deterministic core of the Python prototype's mutation-testing idea, brought into the JS
runtime so the graph-quality claim is checkable in CI: score the impact oracle's
precision/recall/F1 over labeled cases and against the edited-file-only baseline the paper
measured against. A test asserts the oracle out-recalls the baseline on a fixture.

Smaller quality wins

• model_tiers exports PRICING_CURRENCY ("USD") + PRICING_VERIFIED (the date doctor
  checks). The haiku id already matched the canonical dated form, so no id change was needed.
• One shared call-site extractor (src/extract.js) — atlas.js and verify.js each kept a
  copy of the call regex + builtins ignore-list; they now share one module so they can't drift.

On the deeper guard hardening — scoped honestly

The audit flagged protect-paths/secret-redact fail-open without jq. Rewriting their shell
parsing to be safe without jq is fragile to do correctly, so this PR surfaces the risk loudly
(the new doctor jq warning) and defers the full jq-free rewrite rather than shipping a shaky one.

Checklist

• npm test passes (Node 18/20/22) — 222 tests (+4)
• npm run check passes (Biome lint + format)
• New public functions have tests (eval, shared extractor, doctor checks)
• Conventional commit message (feat:)
• CHANGELOG.md updated under ## [Unreleased]
• No new runtime dependency
• Substrate/docs updated (new --json surfaces)

Risk & rollback

• Risk level: low — --json is additive; the doctor checks only report; the shared
  extractor is behaviour-preserving (atlas keeps its own builtins; CALL_RE is identical).
• Rollback plan: revert the commit. No migrations.

Extra checks

• npm run typecheck passes
• Logs contain no secrets/PII

Verification

Full suite + typecheck + Biome green. Manual: forge doctor --json lists the 14 checks incl.
guards exec / jq / atlas / model pricing; forge route "…" --json emits the route object.
eval unit test confirms oracle recall ≥ edited-file-only baseline on a fixture.

Last up: Tranche E (visual docs treatment).

🤖 Generated with Claude Code

https://claude.ai/code/session_01HGZUYFdUb4EViXV5tRid19

---

Generated by Claude Code