feat(openrouter): client-side rate limit for free-tier models only

[mrdulasolutions]

Summary

Fixes user-reported failure:

```
bridge.call(drafts.rerunAgent) failed: sidecar returned error:
OpenRouter HTTP 429: Rate limit exceeded: free-models-per-min.
{"X-RateLimit-Limit":"16","X-RateLimit-Remaining":"0"}
```

OpenRouter caps free-tier models at 16 req/min. The current retry path (3 attempts, exponential backoff capped at 10s) can't outlast a 60-second rate window, and there's no client-side gate — so multiple concurrent agent flows (analyzer, draft generator, rerunAgent, sender lookup) blow through the cap in seconds and the 429 surfaces to the user.

What this PR does

Two complementary changes in `sidecar/src/services/providers/openrouter.ts`, both scoped to `:free` models only (paid OpenRouter models bypass the gate entirely — they don't need it, and threading them through a 16/min bucket would just add latency for no reason):

1. Sliding-window rate limiter

Module-scope timestamp queue tracking the last 16 "issue this call" decisions. Before each free-model fetch:

• Drop timestamps older than 60s.
• If queue is under capacity → push current timestamp, proceed.
• Otherwise → sleep until the oldest timestamp ages out (+50ms cushion to avoid herd retry at the boundary), then re-check.

Awaiters proceed in arrival order because each call awaits before mutating the array.

2. `X-RateLimit-Reset` header parsing on 429

If a 429 slips through (concurrent caller race, sidecar restart resetting the in-process queue, etc.), parse the server's `X-RateLimit-Reset` epoch value and sleep until that point instead of falling back to capped exponential backoff. Caps the sleep at 90s for safety against header bugs / clock skew.

Free vs paid detection

By model-id convention: OpenRouter distributes free models with a `:free` suffix (e.g. `meta-llama/llama-3.3-70b-instruct:free`). The bare model id is the paid tier. `isFreeModel()` is a one-line predicate, easy to extend if OpenRouter changes their naming.

What this PR doesn't do

• Doesn't gate paid models. Per the user's scoping decision.
• Doesn't persist daily limits. OpenRouter also has ~200/day on free; that would need filesystem state and seems premature.
• Doesn't fix Anthropic-side rate limits — different provider, different retry path in `services/anthropic.ts`.

Test plan

• Trigger `drafts.rerunAgent` 20× rapidly with a free model configured — should queue, not 429
• Same with a paid model — should NOT queue (no added latency)
• Inspect sidecar logs for `free-model rate limit reached, queueing` entries during the burst

🤖 Generated with Claude Code