Fix MiniMax M3 effort vocabulary and make agent recovery the single rate-limit retry owner

[usehoplite]

Problem

On the Kilo free gateway, MiniMax M3 (free-2/minimax/minimax-m3:free) accepts high but rate-limits/rejects undocumented xhigh, and when a rate limit did occur the agent surfaced nested 2s/6s/18s/54s delays from the router instead of the intended 10s/20s/30s/40s/60s schedule, continuing through hidden retry layers before stopping.

Changes

Route-true effort vocabulary

• src/llm/model-families.ts: MiniMax M3 now has a model family with its documented effort contract (none/minimal/low/medium/high), so xhigh/max requests clamp to high at request-shaping time on any gateway (Kilo free, nvidia, tokenrouter) while other free models keep the broad gateway vocabulary.
• src/llm/request-plan.ts: the plan compiler resolves the emitted reasoning control once — out-of-vocabulary efforts clamp to the route vocabulary (no-op when the route has no vocabulary), and the cache fingerprint hashes the emitted control.
• src/llm/responses-request.ts: the Responses wire now honors plan-level control suppression (omits the reasoning payload on routes that rejected reasoning controls), matching chat completions; raw preferences no longer bypass suppression. This also corrects the meta muse wire to its documented efforts (medium → high, concise → detailed — two goldens updated).

One retry owner for rate limits

• src/llm/router.ts / src/llm/routing/key-rotation.ts: new retryRateLimits option. The agent turn loop opts out of router-level 429 sleeps (src/agent/turn/loop/round-request.ts), so a 429 on a keyless free route surfaces immediately to the agent recovery schedule instead of nesting 2s/6s/18s/54s waits. Key rotation (one attempt per key), endpoint rotation, provider fallback, non-rate-limit retries, and the effort ladder are unchanged.

Provider reset metadata

• src/llm/wire/response-errors.ts: centralized parsing of x-ratelimit-reset/ratelimit-reset headers and structured error bodies (error.metadata.headers), handling epoch seconds, epoch milliseconds, HTTP dates, and relative seconds; precedence is Retry-After → reset metadata → body hints.
• src/agent/stream-recovery.ts: the recovery planner waits out the longer of the provider reset window and the authoritative 10/20/30/40/60s schedule, capped at 60s, and reports the actual delay.

Effort-ladder regression repair (from a10b7281)

• reasoningWireKey now falls back to the style mapping when the route profile does not govern the wire, restoring distinct rungs for style-mapped providers (reasoning-degradation had regressed to stripping reasoning).
• A ladder with no wire-distinct rung rethrows unattributed failures instead of stripping reasoning and falsely marking the route; attributed rejections still strip after rungs exhaust.

Verification

• npm run typecheck clean.
• Focused new tests: test/llm/minimax-m3-route.test.ts (Kilo wire sends high for xhigh/max, other free models keep xhigh, learned facts override the documented contract), test/llm/router-retry-ownership.test.ts (single-slot 429 rethrows after 1 call, two keys rotate once each, 5xx still retries ×7, default ownership retries ×7), test/llm/reset-metadata.test.ts (10 cases: header/body/epoch/date/precedence), test/llm/responses-control.test.ts, and exact 10/20/30/40/60s + provider-reset assertions in test/stream-recovery.test.ts.
• Broad sweep green: 1,178 tests across 81 files (test/llm, test/admission, test/profiles, test/conformance, plus routing/recovery/provider suites), including the previously regressed test/reasoning-degradation.test.ts (all 26) — repaired in source, without weakening those tests; only the two meta golden snapshots were updated to the documented meta effort contract.