feat: backpressure, search budget and fast ngram comparison

[loris-av]

Summary

Three performance improvements for production workloads:

1. Queue backpressure

• createQueue({maxSize}) rejects new requests when the queue is full (HTTP 503)
• Purges aborted requests before rejecting to reclaim slots
• Configurable via ADDOK_MAX_QUEUE_SIZE env var (default: numNodes * 4)

2. Request timeout

• Requests waiting in queue are rejected after a configurable delay (HTTP 408)
• Configurable via ADDOK_REQUEST_TIMEOUT env var (default: 5000 ms)
• Prevents clients from hanging indefinitely under load

3. Python performance optimizations (bridge.py)

Search budget — Caps the collector pipeline time to prevent slow queries from blocking workers:

• Monkey-patches Search.__call__ and Search.intersect to track elapsed time
• When budget exceeded and results already found, skips remaining intersections
• Configurable via ADDOK_SEARCH_BUDGET_MS env var (default: 200 ms, 0 to disable)
• Logs budget hits to stderr for monitoring

fast_compare_ngrams — Pure Python bigram comparison ~3x faster than ngram.NGram.compare():

• Same algorithm (bigram intersection / union ratio) but avoids re-creating NGram objects
• Replaces addok.helpers.results.compare_ngrams at startup
• Benchmarked on 5K production queries: identical result quality, 3x faster scoring

Test plan

• Lint passes (xo)
• All 56 tests pass (including 8 new queue tests)
• Load test with production-like traffic