Log every stop/nudge intervention (CL-6938)

[TheGreatAxios]

Closes CL-6938. Stacked on #559 (which is stacked on #558) — same files.

This is the gate the project's standing rule points at: no threshold retuning ships before CL-6938 lands.

What it records

One JSONL record per intervention, in the firing worker's own trace dir (interventions.jsonl):

• intervention id + class (stop / nudge / block / clear)
• the trigger's measured value beside its threshold — consecutiveIdentical=5 threshold=5, silenceMs=…, turnsCompleted=…
• provider, model, resolved model family, role (leaf / orchestrator / parent), intent
• run state at the decision: turns used vs budget, total tool calls, read count, edited-path count

Wired at: no-progress, turn-budget, never-acted, never-edited, no-ship, incomplete-report (nudge and stop), report-forced, tool-failure-recovery, stall-nudge, stalled, and the run-loop-owned repetition / deadline / cancelled salvages. Refused parent re-dispatches are recorded from task-tool.ts, since no leaf run exists there to record them.

Writes are fire-and-forget and swallow their own errors — a diagnostic must not be able to fail a run — and the sink defaults to a no-op, so nothing depends on logging being wired.

Reading it back

bun run scripts/intervention-forensics.ts aggregates across local sessions:

• per-intervention counts, split by model family (the axis model-family-policy.ts has no data for today — kimi is a verbatim copy of default)
• measured-value distribution (p50/p90/max) against the thresholds seen
• two false-positive proxies: edited = stops that fired on a run which had already edited files; early = stops that fired before half the turn budget was spent

Neither proxy is proof alone; both are cheap and directional, which is what the tuning history has been missing. It lstats and skips symlinks, so the latest session link cannot double-count.

Verification

bun run check green: 5277 pass / 0 fail. New tests cover the record shape, append ordering, failure-swallowing, and that the director actually emits a nudge and a stop with value ≥ threshold.

Review

needs-sawyer-review — Guardrail 4: touches director/stop-policy internals. Not for auto-merge.

---

Review fix applied: two changes. (a) Relabeled editedWork/earlyBudget in scripts/intervention-forensics.ts and the ARCHITECTURE.md paragraph as neutral context columns, not false-positive proxies — either is equally consistent with a correct stop or a wrong one. (b) Added a real outcome signal: the parent now appends one outcome intervention record per completed dispatch (salvage kind or clean-complete, plus dispatch count), added outcome to InterventionClass, and taught intervention-forensics.ts to report outcome counts by kind. CHANGELOG and ARCHITECTURE updated to describe the outcome records.

[linear-code]

CL-6938