diff --git a/cc-ci-plan/concurrency-restructure-plan.md b/cc-ci-plan/concurrency-restructure-plan.md
new file mode 100644
index 0000000..58d3a3a
--- /dev/null
+++ b/cc-ci-plan/concurrency-restructure-plan.md
@@ -0,0 +1,60 @@
+# Concurrency restructure — concise plan (for review)
+
+Goal: collapse the two concurrency mechanisms (per-recipe flock + pidfile registry/janitor)
+into **one primitive** — a per-app-domain kernel flock — and make lock lifetime bounded by
+construction. Reference: cc-ci `docs/concurrency.md` §8.
+
+## Target invariant chain
+
+```
+lock lifetime ⊆ harness process lifetime ⊆ drone step lifetime ⊆ hard deadline
+```
+
+Never steal a held lock; manage the holder's lifetime. A held lock = live owner (kernel-guaranteed).
+
+## Phases (each independently landable, in order)
+
+### P1 — Lock-lifetime hardening (no behavior change to locking model)
+- Harness startup: `PR_SET_PDEATHSIG(SIGTERM)` + post-prctl `getppid()==1` race check
+  → drone cancel now kills the harness, fixing the leaked-python gap.
+- `.drone.yml` step wrapper: `setsid` + trap killing the process group on TERM/EXIT.
+- Self-deadline: `signal.alarm(90*60)` → teardown + exit. Hard upper bound on any run.
+- Verify: trigger a build, cancel via API, assert harness pid gone + app torn down by next janitor.
+
+### P2 — Flock-probe janitor (replace the pidfile registry)
+- Run start: exclusive flock on `/run/lock/cc-ci-app-<domain>.lock` **before app creation**,
+  held for process lifetime (also serialises double-`!testme` of the same PR/ref).
+- Janitor: probe each candidate's lockfile with `LOCK_EX|LOCK_NB`.
+  Acquirable → orphan → reap **while holding the probe lock** (closes janitor-vs-new-run race).
+  Held → live run → leave; if lockfile mtime > 2× max run time, log loudly (`lslocks` hint) — flag, never steal.
+- Delete: `register_run_app`, `unregister_run_app`, `_run_owner_state`, `ACTIVE_RUN_DIR`,
+  `CCCI_JANITOR_MAX_AGE` + age fallback, pid-reuse guard.
+- Bonus: post-reboot orphans reaped immediately (locks gone = acquirable) instead of after 2h.
+- Keep: `RUN_APP_RE` allowlist, candidate discovery (abra ls + docker-service sweep).
+
+### P3 — Per-run ABRA_DIR (delete the per-recipe flock)
+- Each build: `ABRA_DIR=/var/lib/cc-ci-runs/<n>/abra`, seeded with the server config
+  (verified: abra honors `ABRA_DIR`; fails only on missing `servers/`).
+- `fetch_recipe` becomes a plain clone into a fresh dir (no rm-rf of a shared tree).
+- Delete: `acquire_recipe_lock` + `/run/lock/cc-ci-recipe-*` + the "before fetch_recipe" ordering rule.
+- Wins: same-recipe PRs truly parallel (kills head-of-line blocking §8.2); manual operator
+  activity on `/root/.abra` can no longer corrupt CI runs (§8.3/§8.4 moot).
+- Risk to resolve: orphaned app's `.env` lives in the dead run's tree — janitor must rely on the
+  existing env-less teardown path (docker-service sweep). Confirm `teardown_app(verify=False)`
+  fully works without the `.env`.
+
+### P4 — Config cleanup
+- Drop `concurrency.limit` from `.drone.yml`; `DRONE_RUNNER_CAPACITY` (drone-runner.nix)
+  becomes the single knob (§8.6).
+
+### P5 — Spec rewrite
+- Rewrite `docs/concurrency.md` to the new model (one mechanism, one table) + §10 lock-lifetime chain.
+
+## Out of scope (do not touch)
+- `services_converged()` / paused-is-settled logic; `RUN_APP_RE`; recipe-test gates; warm/canonical apps.
+
+## Open questions for review
+1. P3 server-config seeding: copy vs symlink of `$ABRA_DIR/servers/<server>`? (symlink simpler;
+   copy isolates harder — does abra write into `servers/` during a run? e.g. app `.env` lives there → copy needed, per-run env file then dies with the run dir, see P3 risk.)
+2. Self-deadline value: 90 min ok? (longest observed green run ~35 min incl. flock wait.)
+3. Land P1+P2 together or separately? (P2 testable solo; P1 is pure hardening — could merge.)