Three primitives inside the orchestrator, keeping reviewer isolation intact:

- transcripts/: every spawn's full transcript is preserved in the task folder
  (native: harvested on the completion notification by agent id, sidecar
  fallback by unique description; external: events.jsonl / result.json moved
  out of the spawn temp dir). Gitignored, rescued before any worktree removal.
- scratch/: persistent executable evidence across stages; implementer must
  leave its self-checks there, fixer and QA read them as coverage, not proof.
- log.jsonl: append-only, kinds dead_end and decision only, validated
  structurally after every stage (python or PowerShell, non-blocking).

Access matrix: six stages get the full block, premise-challenge and both plan
reviewers get scratch + log append only. Continuity defined once: retry after
spawn failure and implementer re-run after PLAN_BLOCKED, same profile, native
path only. The "subagents inherit nothing" claim is corrected: they do carry
ambient CLAUDE.md, verified in-harness.

README, maw-context contract, architecture SVG, ACCEPTANCE (three RED rows for
the unverified live mechanisms), PROCESS_METRICS watch condition updated.