feat: pipeline orchestration, normalisation and CSV output

[Hydaspex]

Completes the runnable pipeline: crawl → extract → normalise → persist, with provenance — all driven through the CrawlBackend port, so the whole chain is testable offline with a fake backend. Follows on from #4.

What's included

pipeline/normalise.py — normalise_records: dedupe on identity (region, provider, specialty, metric) (first occurrence wins — callers order sources by trustworthiness) and deterministic sort (region, provider, specialty, metric) with outpatient before treatment. Stable ordering makes run-over-run diffs meaningful. Region is part of identity: the same trust surfaced under two region seeds remains two records.

pipeline/run.py — async run_pipeline(backend, seeds, options):

• Seed = (url, region); the region is threaded into every record discovered under that URL
• CrawlRun provenance captured automatically (backend from the concrete class name, uuid4 run_id, tz-aware started_at)
• frozen PipelineResult(run, records)
• extractor failures surface as absent records (its contract); backend failures propagate loudly

io/csv_handler.py — write_records_csv with the fixed COLUMNS contract (versioned with the golden schema): None → empty cell, empty input → header-only file, parent dirs created.

cli.py — nhs-scraper entry point: --backend, repeatable --seed URL=REGION, --output, --limit, --max-depth; backend constructed lazily so importing the CLI never pulls crawl dependencies.

Tests (24 new)

• test_normalise.py — first-wins dedupe, distinct metrics kept, multi-region identity regression, sort order, empty input
• test_pipeline_run.py — keystone: the full crawl → extract → normalise chain over the characterisation fixture reproduces the golden dataset exactly, plus provenance, region threading, multi-seed aggregation, empty-seed rejection, empty crawl
• test_csv_handler.py — column contract, None → empty, header-only, nested dirs
• test_cli.py — seed parsing (incl. = in values), arg defaults, repeated seeds, unknown/missing-backend errors

pyproject.toml gains [project.scripts] nhs-scraper.

Verification

Pipeline (golden-order output, dedupe/sort, region threading, edge cases) and CSV handler (column order, None handling, empty case) both exercised locally in the sandbox — all green.

Design decisions to review

1. Identity includes region (corrected) — the initial commit keyed dedupe on (provider, specialty, metric); test_multiple_seeds_aggregate caught identical trust records under two region seeds collapsing 8 → 4. Identity is now (region, provider, specialty, metric), with a regression test. Duplicates only collapse within identical identity.
2. Backend name from type(backend).__name__ — zero-config provenance; alternatives (explicit name field) add plumbing for little gain.
3. CLI defaults to a single example seed — deliberate: running bare nhs-scraper is a smoke test, not a full crawl.