feat: transient-error retry policy with failure telemetry

[Hydaspex]

Closes the last gap from the drift/retries review: previously there were no retries anywhere, and per-page crawl failures vanished silently. Follows on from #7.

What's included

ports.py — RetryPolicy (frozen, validated)

• attempts >= 1, backoff_seconds >= 0; linear backoff (backoff * attempt) — deliberate politeness towards a public NHS site, where exponential retries are the wrong posture
• Docstring encodes the rule: validation failures are never retried — only transient transport errors and soft (unsuccessful-result) failures

backends/crawl4ai_backend.py — retry engine

• _arun_with_retries: retries arun on raised exceptions (transport/browser errors) and unsuccessful results (soft failures, e.g. timeouts)
• Exhaustion semantics: soft failures → CrawlError with the site's error message; transport errors → the original exception propagates
• No policy = single attempt — previous behaviour preserved exactly
• sleeper injection hook so tests record backoff delays without sleeping

Failure telemetry

• backend.last_failed_pages — URLs of unsuccessful page results, recorded per crawl and reset between calls; pages are still skipped (partial results kept) but no longer silently
• PipelineResult.failed_pages — run_pipeline aggregates telemetry via duck-typed getattr(backend, "last_failed_pages", ()), so backends without the attribute are unaffected
• CLI prints a failed-pages warning after the summary

CLI — --attempts (default 3) / --backoff (default 2.0) wired through build_backend.

Tests (17 new + 2 adjusted) — policy validation; 6 scrape-retry scenarios (transient-then-success, soft-fail-then-success, exhaustion both modes, single-attempt default, recorded backoff [2.0, 4.0]); crawl telemetry + reset; pipeline aggregation with partial results kept; no-telemetry backend; CLI flags. The two test_preflight.py monkeypatch lambdas were widened to *a, **kw for the new build_backend signature.

Verification

8 scenarios executed in the sandbox before pushing — transient retry, soft-fail retry, exhaustion both modes, single-attempt default, telemetry, linear backoff, validation — all green.

Design decisions to review

1. Linear over exponential backoff — politeness posture for a public NHS site; exponential is available later via a strategy field if ever needed.
2. Soft failures retried, validation failures never — a success=False from the crawler is usually transient; a ValueError from our own validation is deterministic.
3. Duck-typed telemetry (getattr convention) over a port change — keeps the protocol minimal; a formal TelemetryBackend protocol is a possible later refinement.