Tasks: generalize-adaptive-collection-governor

1. Make adaptive collection the default in the shared governor

• 1.1 Add shared constants DEFAULT_PACING_INITIAL_INTERVAL_MS = 1000 (slow-start discovery seed) and DEFAULT_PACING_MIN_INTERVAL_MS = 250 (the one owner number — the rate ceiling), exported from connector-http-governor.ts.
• 1.2 Default pacingInitialIntervalMs / pacingMinIntervalMs to those constants so a bare createConnectorHttpGovernor({ name }) yields adaptive pacing; keep pacingInitialIntervalMs: 0 as the explicit opt-out (no-pacing, byte-identical pre-convergence path, snapshot() === null).
• 1.3 Add tests proving the bare governor cold-starts at the discovery seed, ACCELERATES under sustained success toward the ceiling, BACKS OFF on a throttle (legible in the snapshot), and never crosses the ceiling.

2. Warm-start runtime seam (persist learned rate across runs)

• 2.1 Add the restoredIntervalMs option to createConnectorHttpGovernor (seed ProviderPacing warm-started, clamped to the ceiling).
• 2.2 Expose snapshot(): PacingSnapshot | null on the returned governor.
• 2.3 Add shared helpers readPersistedPacingInterval(stateSlice, opts) (restore + staleness guard), buildPacingStateFields(governor, opts) (persist), with default state keys + a derived staleness window.
• 2.4 Tests: warm-start round-trip (a run persists; the next restores), stale-discard, absent/malformed → cold start.

3. Six API connectors inherit adaptive behavior with no per-connector rate code

• 3.1 Confirm the adoption smoke (github/ynab/notion/strava/oura/spotify) still throws <name>_rate_limited byte-identically under default-on pacing.
• 3.2 Confirm each connector's full suite stays green and FAST (instant-timer the one suite — github index — that exercises the real gh() pacing path).
• 3.3 Wire github as the reference warm-start adoption: re-seed the governor from state.user at run start; persist the FINAL learned interval onto the declared user stream cursor at run end (a real cursor — the runtime gates STATE on declared streams, so warm-start state never rides a synthetic stream).

4. collection_rate observability for all governor-using connectors

• 4.1 Add buildCollectionRateProgress(governor) returning the CollectionRateProgress shape (carries only rate numbers + last back-off reason, no account content), or null when pacing is opted out (honest absence).
• 4.2 Emit collection_rate from github via the shared helper; it flows through the existing runtime→spine→console path with no new UI.
• 4.3 Test: the bare governor's snapshot yields legible rate state with no account-content leakage; opting out yields null (no false zero).

5. New-connector experience documented

• 5.1 Document in the governor module header: "to add an API connector with fastest-safe adaptive collection, call createConnectorHttpGovernor({ name }) — discovery, warm-start, back-off, and observability are automatic," and make that true.
• 5.2 Note in the module header that browser-bound connectors and reddit are Phase B and do NOT use this factory.

6. Gates

• 6.1 chatgpt suite green (controller path untouched; 155/155).
• 6.2 All six API connector suites green (144/144).
• 6.3 governor / pacing / budget / send-governor / adoption suites green, with the adaptive-parity tests proving accelerate-under-success.
• 6.4 tsc --noEmit clean; biome/ultracite clean on changed files.
• 6.5 openspec change validates --strict.

7. ProviderProfile: the safety/pressure quantity is a REQUIRED declaration (spec §3, §9-C5)

Closes the §3/§9-C5 architectural gap: §1.2 above hoisted ChatGPT's pacingMinIntervalMs=250 as a shared default; the spec's bar is the opposite — "no shared default for a safety- or pressure-shaped quantity; a missing field is a BUILD ERROR, not a silent borrow of ChatGPT's number." This section makes the pacing ceiling a REQUIRED per-provider declaration and SUPERSEDES the default-fallback in §1.2. (The terminal-gap maxRecoveryAttempts and cooldown maxCooldownCycles slices were already lifted into required per-provider profiles in commit 753012ab — §10-A/§10-B; this section completes the set by doing the same for the pacing ceiling.)

• 7.1 Add packages/polyfill-connectors/src/provider-profile.ts — the ONE declared home for the profile field set: ProviderPacingProfile (pacingMinIntervalMs, compile-time required on the governor), ProviderTerminalGapProfile (maxRecoveryAttempts), ProviderCooldownProfile (maxCooldownCycles), and the union ProviderProfile. These are the safety-/pressure-/terminal-shaped quantities (getting them wrong off the wrong provider's numbers risks a ban, a stall, or a dishonest health verdict); discovery seeds / AIMD horizons stay derived defaults, NOT forced here.
• 7.2 Make the pacing ceiling REQUIRED at the type level: replace the optional pacingMinIntervalMs?: number on ConnectorHttpGovernorOptions with a non-optional profile: ProviderPacingProfile. A bare createConnectorHttpGovernor({ name }) is now a tsc error (the spec's build-error bar). Add a loud startup throw as the JS-caller backstop (no silent shared default). DEFAULT_PACING_MIN_INTERVAL_MS survives only as a NAMED, AUDITED reference for tests — the governor no longer falls back to it.
• 7.3 ChatGPT is unchanged (pure refactor): ChatGPT runs on ProviderBudgetController, not this factory, and keeps CHATGPT_DEFAULT_PACING_MIN_INTERVAL_MS = 250. Terminal maxRecoveryAttempts=3 and cooldown maxCooldownCycles=8 are unchanged. No live-behavior change.
• 7.4 The six governor-using connectors (github/notion/oura/spotify/strava/ ynab) each declare unauditedConservativePacingProfile() (a DELIBERATE 1000ms conservative placeholder, ~4× slower than ChatGPT's audited 250ms — NOT a borrow). Terminal/cooldown slices are NOT declared for them: they opt OUT of those loops (terminalization skipped via terminalGapProfileForConnector → null; cooldown escalation disabled) rather than inheriting ChatGPT's budgets.
• 7.5 Conformance test src/provider-profile-conformance.test.ts PINS the build-error bar: a @ts-expect-error on the bare governor call (unused suppression ⇒ tsc fails if the field is ever made optional again), the runtime backstop throw, a static source scan proving every governor-using connector declares a profile and never hard-codes 250, and that the conservative placeholder is slower than ChatGPT's number. Verified to FAIL when a connector's profile is removed (both the conformance test and tsc go red).
• 7.6 Gates: tsc --noEmit clean in polyfill-connectors (RI clean except the pre-existing unrelated google-data-portability.ts error); full polyfill-connectors suite 2086/2086; the 5 SLVP suites green (terminal-gap 18, cooldown-recovery-eligibility 4, escalation-l5 12, push-escalation-l8 5, provider-pacing 35); ultracite clean on all changed files.

7b. OUT OF SCOPE here — per-connector behavioral audit (the §9-C5 follow-up)

• 7b.1 Audit each of the six governor-using connectors' REAL 429 semantics (is a bare 429 whole-account-hot? fast-open attempts?) and binding-quota axis, and replace unauditedConservativePacingProfile() with that provider's observed ceiling. Tracks the spec §3 pressureSignal / servedBackoffCostMs fields.
• 7b.2 Where a provider warrants it, declare its terminal-gap maxRecoveryAttempts and cooldown maxCooldownCycles from observed non-transient -error and recovery-window behavior (today only ChatGPT declares these).
• 7b.3 Each new/updated profile validated by a supervised live run (owner-run; no live calibration in the interface change).