fix(smokes): restore the public smoke fleet and the capability roster regression

[songoow]

Behavior

Full Public Smokes has been red on main for two days (no successful run in the last ~100), with 20+ scripts failing per shard sweep. This restores the fleet by fixing one production regression and realigning the smoke contracts that drifted behind merged behavior changes. Locally 9 of the 12 first-failing scripts pass after this PR; the remaining three are listed under Boundaries.

Production fix

resolve_agent_capabilities re-read the goal registry file to obtain the registered-agent roster before the capability-memory observation. A relative or unreadable registry path silently emptied that roster, the typed handler rejected the read-only observation as an unregistered agent, and quota should-run failed with EffectRuntimeInternalError instead of returning a projection. The resolved identity already carries the authoritative roster, so it is passed through; roster entries keep agent-id normalization (they are not capability tokens).

Contract realignments

• fix(heartbeat): recover unsettled host turns #4302 closeout wording: install-local, blocker-push and heartbeat-prompt asserted the pre-fix(heartbeat): recover unsettled host turns #4302 guard receipt; 2 stalls->replan line.
• fix(heartbeat): reserve readable Reward Memory prompt headroom #4333 Reward Memory headroom: the heartbeat fixtures measured prompts against the bare INTERFACE_BUDGET_CHARS ceiling; production budgets add the 640-char outcome-contract headroom when the marker is present. The fixtures now mirror build_interface_budget, and in-process comparisons against CLI output opt out of the Reward Memory contract explicitly, because the CLI derives it from the goal registry (absent means off).
• refactor(quota): unify typed capability resolution and repair bindings #4254 typed capability owner: capability-gate-projection imported three private helpers that moved into capability_gate.ts; it now reads the public gate. The requirement-ordering case now correctly expects no gate when no candidate declares a requirement.
• refactor(authority): unify Todo presentation and provider conformance #4283 presentation ordering: todo-user-gate-readmodel pinned a display order that native records without a v0 wire coordinate no longer have; the lane membership assertion is order-independent.
• fix(quota): preserve action and monitor receipt identities #4335 monitor identity: the rejection report names the actually selected monitor and its declared material-change flag instead of an empty identity.
• fix(loopx): reconcile skill discovery and persistent heartbeat startup #4296 docs/skill wording and the thin agent-input projection: assertions follow the current two-layer prompt documentation and read the embedded guard command from the task body.

Module budgets and classification

agent-capabilities and handoff were registered without a manual/help-only classification; both are command-specific surfaces and join the help-only set. Three cohesive owners are extracted to return their modules under the 1000-line budget: quota_failure_report.py (honest failure/validation payloads, moved verbatim), support_control_chat.py (chat + dashboard registration), project_lifecycle_reward.py (reward overlay registration). quota.py 1118→928, support_control.py 1015→889, project_lifecycle.py 1014→963.

Validation

• Focused suites: test_quota_settlement_cli 57 passed, heartbeat prompt suites 44 passed, capability-memory CLI tests 3 passed (run via importlib against the repo checkout), shadow/lease suites unaffected.
• Full TypeScript control-plane suite: 1,490 passed (the PostgreSQL integration test requires a local service and fails identically on an unmodified checkout).
• Ruff on all touched files; the extractions also remove the imports that only the moved code used (33→17 pre-existing warnings on those modules; the remainder are pre-existing broad-exception handlers).

Boundaries

• turn.py (1,114 lines) exceeds the budget on main before this PR; it is a single large handler and pin-vs-extract is a maintainer decision this PR does not take.
• configure-goal-smoke needs a complete reward_memory_experiment_config_v1 fixture since the enablement preflight landed; that is a corpus-contract decision, not an assertion drift.
• monitor-poll-writeback's capability-declaration parity case fails identically on unmodified main and is not addressed here.

[songoow]

Closing as superseded: #4355 landed the same fleet restoration while this PR was in flight, covering the smoke contract realignments, the agent-capabilities/handoff classification, and the capability-memory failure on quota should-run. I verified on c5e43a161 that the previously failing scripts here now pass.

Two notes for the record:

• fix: restore public smoke contract parity #4355 resolves the capability-memory failure by failing open (except Exception: state = {}), which also drops remembered observations whenever the status packet's registry path is unreadable. This PR instead passed the already-resolved roster from the agent identity through to the observation, keeping remembered capabilities live in that case. If that distinction matters, I can reopen that one change as a small follow-up on top of the fail-open guard.
• fix: restore public smoke contract parity #4355 pins quota.py (1118) and support_control.py (1015) at their baselines "while each cohesive extraction lands". This branch contains two such extractions (the honest quota failure/validation payload owner, and the chat/dashboard registration pair) that retire both pins; I will resubmit them as a focused follow-up against the new baseline.