Tasks: ship-adaptive-collection-rate-controller

1. Delete the launch-jitter floor as a rate floor

• 1.1 Lower CONVO_DETAIL_PAUSE_MIN_MS/CONVO_DETAIL_PAUSE_MAX_MS defaults from 1500/3000 to an ε anti-phase-lock window (0/150).
• 1.2 Update CHATGPT_SERIAL_TUNING and the frozen-default comments so the GCRA hint is the sole rate authority and jitter is ε noise only.
• 1.3 Rewrite the existing "lane retains its 1500ms launch delay" test to assert the lane's launch wait is now governed by the pacing hint, with jitter bounded to ε.

2. Slow-start discovery + warm-start

• 2.1 ProviderPacing: accept restoredIntervalMs (seed _currentIntervalMs, clamped to never be faster than minIntervalMs) and add snapshot().
• 2.2 ProviderBudgetController: re-expose snapshotPacing().
• 2.3 Lower the ChatGPT cold-start initialIntervalMs default from 2500 to 1000 (discovery seed, used only without fresh learned state).
• 2.4 Connector: read the persisted learned interval from durable state at run start, apply the staleness guard, pass it as restoredIntervalMs.
• 2.5 Connector: stamp the controller's final learned interval to durable state at the end of the detail pass.
• 2.6 Tests: warm-start restores near the prior run's interval; a stale resume falls back to the cold seed; restore never crosses the ceiling.

3. The one owner number: the rate ceiling

• 3.1 Confirm PDPP_CHATGPT_PACING_MIN_INTERVAL_MS is the single ceiling knob with a safe default; document it as the only number.
• 3.2 Neutralize/annotate the concurrency-AIMD as inert under maxConcurrency === 1 (no second control dimension).
• 3.3 Test: additive increase floors at the ceiling and never crosses it.

4. Drain-within-budget recovery

• 4.1 Make recoverPendingMessageDetailGapsBeforeForwardRun void-returning; remove the return on stoppedWithPending in runConversationsAndMessagesStreams.
• 4.2 Continue to the forward walk; return only when deps.runBudget?.shouldStop().
• 4.3 Reconcile the tailStop abort: now a micro-optimization, not required.
• 4.4 Test: a source-pressure recovery stop with remaining budget proceeds to the forward walk, advances the cursor, and keeps gaps durable; genuine budget exhaustion still defers.

5. Observability

• 5.1 Emit a collection_rate structured progress event (current interval, effective rate, ceiling, last back-off + reason) on speed-up/back-off transitions.
• 5.2 Add a "Collection rate" readout to the connection-detail diagnostics region, degrading to an explicit unknown.
• 5.3 Test: the event carries no account content; the console readout degrades honestly when state is absent.

6. Simulation tests proving the ideal

• 6.1 Under sustained success the effective interval DECREASES toward the ceiling over time (throughput rises) — the test that catches the flat-19/min posture.
• 6.2 On injected 429/pressure the interval multiplicatively increases and recovers.
• 6.3 The interval NEVER crosses the ceiling.
• 6.4 Warm-start: a run resumes near the prior run's learned interval, not the cold default.
• 6.5 Drain-within-budget: an open pressure circuit with remaining budget does NOT terminate; advances the cursor; defers gaps durably.

7. Gates

• 7.1 Full chatgpt suite + governor/pacing/run-budget + reason-discrimination suites green.
• 7.2 tsc + lint baselines.
• 7.3 openspec validate --all --strict green.
• 7.4 Owner: supervised live calibration run (NOT done in this change).

8. In-pass circuit-continue (the in-pass analogue of §4)

Closes the live run_1781150455121 defect: a circuit_open provider-budget gate DURING a detail pass was bucketed with the genuine run caps (chatGptRunCapReasonFromProviderGate), so a transient upstream-pressure circuit deferred the entire remaining tail and finished the run after 136 s of a 900 s budget. §4 fixed the recovery→forward-walk hand-off; this fixes the in-pass case.

• 8.1 CircuitBreaker.remainingCooldownMs() + ProviderBudgetController.circuitCooldownMs() expose the exact open→half_open cool-down so a transient back-off can sleep it precisely.
• 8.2 RunBudget.remainingWallClockMs() (+ ChatGptRunBudget delegate) bounds the cool-down wait by the run's true remaining wall-clock budget.
• 8.3 fetchConversationDetailWaitingOutCircuit: on a circuit_open planned defer, wait the cool-down (bounded by remaining budget), then retry the same conversation; genuine budget exhaustion (max_wall_clock/max_detail_fetches) propagates immediately to the run-cap tail path. Forward-progress guard: a bounded cycle count + clock advance guarantees a never-closing circuit converges to a durable defer (no infinite loop).
• 8.4 Tests: an open circuit with budget remaining CONTINUES (waits, resumes, collects the full tail); genuine wall-clock exhaustion during a wait still defers; a never-closing circuit is bounded; the live 136 s/900 s shape now runs the full budget. Src tests for the three new primitives. All green; tsc 0; biome clean.