docs(docs): V1-2.4 maquettes proposal, verification worker risk register

docs/design/v1-2-4: twelve functional mockups on the existing identity
(catalogue, editor, narrow read-only, verification states, review, admin
navigation, FR/EN labels), not validated. V1-2 contract §13.1: accepted risks
VR-1 (CPU contention, MEDIUM) to VR-5 with their re-examination condition.
This commit is contained in:
bchanot committed 2026-11-02 00:11:04 +01:00
1 parent 8db62b43e2
commit f00d8b4e4c
17 files changed
+688

No files matched your search

@@ -1806,6 +1806,20 @@ Residual risks:
readers refuse bigger snapshots;
- RR-2 stays open (§16.12).
### 13.1 Accepted risks of the verification worker (register, V1-2.3)
Owner decisions of 2026-11-02 (BDR-093): accepted for the sandbox scope, not
fixed, and no authorisation to deploy. Each line is re-examined before
`worker-verify` runs on DEV or verification reaches external users.
| Id | Severity | Scenario | Existing limits | Effect on attempts and resources | Re-examination |
|---|---|---|---|---|---|
| VR-1 | MEDIUM | A document whose verdict needs about 60 s of CPU, on a host where the child gets less than one full core (the VPS shares its cores with the Supabase stack; `--cpus 1` is a ceiling, not a guarantee), reaches the 90 s wall deadline before the 60 s CPU budget: the parent kills it, the end reads `Timeout`, reported `verification.worker.timeout`, a transient code. Reproduce: a document at the CPU budget on a loaded host. | child RLIMIT_AS 1 GiB, CPU 60 s (SIGXCPU, definitive), wall 90 s; one child at a time per worker; five attempts; delays 2, 5, 10, 20 s | up to five runs of about 90 s plus 37 s of delays, about 8 min of worker time for one round, instead of one run for a definitive close; other rounds wait behind it on a single worker; no effect on publication (the round closes `unavailable` at the limit) | before `worker-verify` on DEV or external users: measure the child's CPU time at the end (`wait4` rusage, an `unsafe` exception the owner deferred) or reserve a core for the worker |
| VR-2 | LOW | A frame header over 64 KiB (several hundred drawings named) is reported `resource_limit` instead of `caps` | definitive at first occurrence | one run | when the editor screens show worker codes (V1-2.7) |
| VR-3 | LOW | An object larger than the policy is passed without bytes: failures `[caps, drawing_missing]`, never `size_mismatch` | bucket limit 2 MiB prevents it in practice | none (verdict failed at once) | same |
| VR-4 | LOW | SIGABRT also covers a child panic and a host ENOMEM, both closed `resource_limit` | panic deterministic on a frozen input | one run; a host ENOMEM needs a resubmission | before DEV activation |
| VR-5 | LOW | A verdict bound to another round, sent with a valid token, answers `stale` and waits for the lease | only after a core or worker regression | up to five leases of 10 min before `lease_expired` closes it | before DEV activation |
---
## 14. Equine sheet