feat(autodev): autodev.toml schema reader + parameterised hook prompt

[bahadirarda]

Summary

Phase 2-4 of the autodev hook redesign. Builds on PR #16 (cwd guard) by adding the schema reader for <project>/.clawtool/autodev.toml, parameterising the continuation prompt, and adding the enabled = false operator-pause path.

Pipeline order in runAutodevHook

1. cwd guard      — exit 0 silently when no marker resolves upward from $PWD
2. config load    — parse the marker; fail-soft to defaults on read/parse error
3. enabled guard  — exit 0 silently when `enabled = false`
4. arm-flag guard — exit 0 silently when `autodev.enabled` flag absent
5. cap guard      — emit cap-reached message + exit when counter ≥ cap
6. fire           — increment counter, emit parameterised prompt

Schema

name = "myproj"                     # `Project: myproj` line in prompt
enabled = true                      # nil → true; &false → operator-pause
cap_per_session = 200               # overrides global 200 cap; non-positive → 200

[loop]
steps = [
    "step one",
    "step two",
]
stop_command = "/myproj-stop"       # empty → /clawtool-autodev-stop default

Backwards-compat

• A minimal marker (e.g. name = "..." alone) parses cleanly. Empty [loop] steps → historical 5-step clawtool fallback. No operator action required for existing armed projects.
• The .clawtool/autodev.toml shipped in PR feat(autodev): cwd guard — .clawtool/autodev.toml as project opt-in marker #16 is fully schema-compliant; verified by TestRunAutodevHook_RealClawtoolMarker_ParsesCleanly (regression guard against schema breakage).
• Fail-soft parser: malformed TOML returns zero-config rather than blocking the hook. A Stop-hook fires every turn-end; a parse error must not silently disarm the loop.

Files

• internal/cli/autodev_config.go (+105 lines) — autodevConfig struct, loadAutodevConfig, the three effective* helpers
• internal/cli/autodev.go (+94/-17 lines) — pipeline ordering in runAutodevHook, autodevHookPrompt(triggerN, cfg) parameterisation, renderAutodevLoopSteps helper with historical fallback
• internal/cli/autodev_config_test.go (+330 lines) — 10 new tests covering schema parse paths, fail-soft cases, prompt rendering, the live clawtool marker

Verification

gofmt -l                           # clean
go vet ./internal/cli/             # clean
go test -race -count=1 -timeout=60s -run \
  "TestRunAutodevHook|TestFindAutodevTOML|TestLoadAutodevConfig|TestAutodevHookPrompt" \
  ./internal/cli/ -v               # 15/15 PASS
go test -race -count=1 -timeout=180s ./internal/cli/   # ok 14.643s

What this unlocks

• Operators can now author per-project loop steps without a Go PR. Edit .clawtool/autodev.toml, save, next turn-end picks up the new template.
• enabled = false lets an operator pause the loop in a specific project without removing the marker (which would also drop forward-compat schema fields).
• Cap-per-session lets long-running refactors raise the budget for a particular project, and short-running utilities lower it.
• ADR-039 Phase A.5 (the --include-features ideator flag) now has a place to plug into: [loop] steps can reference clawtool ideate --include-features once that source-tier ships.

Wiki context

ADR-037 (wiki/decisions/037-autodev-hook-redesign.md, gitignored). This PR ships axes 2-4 of the four-axis design; the firing pipeline in §3 is now in place. ADR-037 §3 axes still pending: quiet-state gate (§3a), substrate probe code/wiki/stalled mode selection (§3b), autodev.lock JSON staleness rules (§3c). Those land in follow-up PRs.

Hygiene

Commit ran via mcp__clawtool__Commit (Conventional Commits + pre_commit rules.toml gate). gofmt-clean and race-clean rules trigger on this changeset (Go files touched) — both verified clean above. No internal-doc-id leakage in user-facing surfaces (CLI help, command help, prompt template). The [[037 Autodev hook redesign]] reference in the in-source comment is intentional — it's a Go source comment, not a user-facing surface, and the ADR is the canonical design record for this code.