feat: state worker (builtin migration)

.github/workflows/create-tag.yml

@@ -38,6 +38,7 @@ on:
           - slack
           - telegram-bot
           - shell
+          - state
           - worktree
           - storage
           - web

.github/workflows/release.yml

@@ -32,6 +32,7 @@ on:
       - 'slack/v*'
       - 'telegram-bot/v*'
       - 'shell/v*'
+      - 'state/v*'
       - 'storage/v*'
       - 'web/v*'
       - 'worktree/v*'

docs/adr/0001-state-worker-store-location.md

@@ -0,0 +1,51 @@
+# ADR 0001: State worker — store location
+
+## Status
+Accepted (2026-07-06).
+
+## Context
+The engine builtin `iii-state` is not just a trigger handler: it owns a key-value
+store that lives in the engine process (`engine/src/builtins/kv.rs`,
+`BuiltinKvStore`, in_memory or file_based with a background flush loop).
+Migrating `iii-state` to a standalone worker forces a decision about where that
+store lives. Coupling analysis on the engine found ZERO inbound engine-internal
+consumers of `state::*` (everything reaches state over the function bus), so the
+functions can move; the question is only the store.
+
+## Options
+
+### (a) Store lives in the worker process — RECOMMENDED
+Port `BuiltinKvStore` (file_based/in_memory, identical on-disk format) into the
+worker; `redis` adapter ported as-is for multi-instance/production.
+
+- Durability: store dies with the WORKER process instead of the engine process.
+  `file_based` keeps the same crash window (unflushed writes within one
+  `save_interval_ms` cadence); `redis` is unchanged. An engine restart no longer
+  wipes in_memory state (the worker survives it); a worker restart now does.
+- Latency: `state::get/set` change from in-process calls (µs) to engine⇄worker
+  WebSocket round-trips (ms). This is the price of every builtin migration but
+  state is the hottest path — MUST be benchmarked (see Acceptance criteria).
+- No engine changes required; ships entirely from the workers repo.
+
+### (b) Engine exposes the store as a service; worker is a thin façade
+Engine keeps `BuiltinKvStore` behind new internal functions; the worker proxies.
+- Keeps in-engine durability but ADDS a hop to every call, and requires an
+  engine-repo change — out of scope for the workers repo. Rejected unless (a)'s
+  benchmark fails the parity gate.
+
+### (c) State remains a builtin
+No migration. Rejected as the default: contradicts the registry-migration
+program, but it is the fallback if the latency benchmark shows unacceptable
+regression for real workloads (e.g. Motia per-step state reads).
+
+## Decision
+Option (a). Tasks 2–13 of the state-worker plan are written against it.
+
+## Acceptance criteria
+1. Durability profile documented in the worker README (parity table): what is
+   lost on worker crash per adapter (`kv in_memory`: everything; `kv
+   file_based`: ≤ one save_interval window; `redis`: nothing).
+2. Latency delta acknowledged: `state::get/set` move from in-process calls
+   (µs) to engine⇄worker WS round-trips (low ms). The formal p50/p95
+   benchmark gate was waived by project decision on 2026-07-06; the README
+   parity table records the order-of-magnitude delta instead.