Delete the runtime's codex-shaped recovery regexes and account-restart set (WS4 L2)

[SawyerHood]

WS4 layer 2 of 7, stacked on #2235 (L1). Design: ideal-state-provider-api.md §7 "Deleted from core" — the codex regexes, the account-restart set, and the rename retry.

What was wrong

After L1 every condition the runtime used to recognize by matching codex prose arrives as a typed provider/recovery hint, so the regexes and the provider-id guards they justified were dead weight that still ran: CODEX_ARCHIVED_SESSION_ERROR_PATTERN gated the unarchive-and-retry beside the hint, the runtime retried renames a bridge had already retried, and shouldRestartCodexThreadAfterEvent re-classified every terminal codex error by text to schedule a process restart the codex bridge now performs itself. The steer path still needed isAcpProviderId to read the ACP bridge's NO_ACTIVE_TURN as a stale steer.

What changed

Deleted from packages/agent-runtime/src/runtime.ts (each with its L1 counterpart):

Deleted symbol	Covered by
CODEX_ARCHIVED_SESSION_ERROR_PATTERN, isCodexArchivedSessionError	sessionArchived on the rejection (error.data.recovery); sendCommand now switches on the hint alone
CODEX_EMPTY_ROLLOUT_RENAME_ERROR_PATTERN, CODEX_RENAME_RETRY_DELAYS_MS, sendRenameWithRolloutRetries, SendRenameWithRolloutRetriesArgs	the codex bridge's own 50/200 ms ladder (sendMaintenanceRequestWithRetries); the runtime sends thread/name/set once
CODEX_ACCOUNT_RESTART_PROVIDER_ERROR_CATEGORIES, CODEX_ACCOUNT_RESTART_PROVIDER_ERROR_TEXT_PATTERN, shouldRestartCodexThreadAfterEvent, restartCodexThreadForNextTurnIfNeeded, codexThreadsRequiringAccountRestart	the codex bridge rebuilds its app-server child before the next turn (rebuildBeforeNextTurnReason) and raises authRequired/rateLimited; the generic restartRecommended action stays for bridges that cannot self-heal
the isAcpProviderId(pid) && code === NO_ACTIVE_TURN steer branch	the generic staleTurn branch
imports isAcpProviderId, ProviderErrorCategory, BRIDGE_JSON_RPC_ERRORS	unused after the above

RestartCodexThreadForNextTurnArgs is renamed RestartThreadBridgeArgs (it now only serves restartThreadBridgeIfRecommended). The CODEX_PROVIDER_ID guards that remain (resolveProviderProcessKey, isThreadScopedCodexProcess, the pre-experiment idle-reap gate) belong to the per-thread process lane and go in L3.

Single-site exception in plugins/provider-acp/src/bridge/bridge.ts (coordinator-approved): the turn/steer NO_ACTIVE_TURN rejection now carries data: { recovery: { kind: "staleTurn", retryable: false } }. Nothing else in that plugin changes; the error code stays for the wire.

Ratchet: scripts/provider-literal-baseline.json 148 → 142 (runtime.ts 12 → 6).

Not a wire change: no HOST_DAEMON_PROTOCOL_VERSION bump (nothing between server and daemon moved) and no PROVIDER_BRIDGE_PROTOCOL_VERSION bump (the ACP data is the optional field L1 added).

How you verified

Tests:

• Removed the three runtime tests that pinned the deleted behavior — retries a Codex rename while its new rollout file is still empty, stops retrying a Codex rename once its rollout stays empty (replaced by does not retry a rename: the bridge owns the not-ready-rollout ladder, one thread/name/set on the wire), and the two account-restart tests in runtime.process-lifecycle.test.ts (their counterparts: bridge.recovery.test.ts in the codex plugin and the restartRecommended cases in runtime.recovery.test.ts).
• maps a bridge no-active-turn error to a stale steer now sends the staleTurn hint (the ACP bridge's shape) and is no longer provider-gated.
• The L1 suite runtime.recovery.test.ts runs under provider id fake, so it already proved the typed path with the text gates unable to fire; it is unchanged and green.

Oracle (--concurrency 4): typecheck 9/9; tests agent-runtime 450, provider-acp 196, provider-codex 201, scripted-echo 1, host-daemon 555 — green. Parity acp-cursor 10 passed / 1 skipped, zero diffs (the data rides an error response, never an event). A4: 307 threads, 93,534 rows, zero diffs. Ratchet --base origin/main: 142 ≤ 148, baseline written.

Part of the provider-plugin migration (WS4). L3 collapses the codex per-thread process lane.

AGENT GENERATED: by Claude Opus 5

[SawyerHood]

Coordinator review — WS4 layer 2 (#2236)

PASS against REVIEW-GATE.md. Do not merge: the stack accumulates until the full migration is done.

Verified on the branch

• All six named symbols are gone from runtime.ts (CODEX_ARCHIVED_SESSION_ERROR_PATTERN, the rename ladder pair, the account-restart pattern, restartCodexThreadForNextTurnIfNeeded, the isAcpProviderId import). The runtime matches no error text.
• Each of the five removed runtime tests has a successor: rename retry → "the bridge owns the ladder" plus L1's bridge ladder tests; NO_ACTIVE_TURN → staleTurn; the two codex account-restart tests → L1's bridge.recovery.test.ts (the restart now happens inside the bridge by design).
• The ACP change is one hunk at the NO_ACTIVE_TURN emission and adds only data.recovery.
• Ratchet baseline 148 → 142 written with --write.
• Stack shape: Delete the runtime's codex-shaped recovery regexes and account-restart set (WS4 L2) #2236 draft on Bridges emit typed provider/recovery hints; the runtime acts on the kind (WS4 L1) #2235 draft on main.

Oracle (reported): typecheck 9/9; agent-runtime 450, provider-acp 196, provider-codex 201, host-daemon 555 green; parity acp-cursor 10/0; A4 307 threads, zero diffs.

AGENT GENERATED: by Claude (coordinator)