feat(benchmarks): add end-to-end retrieval evaluation harness

[AhmedIrfan7]

What

Add a reproducible end-to-end evaluation harness that measures VidXP's complete public retrieval path in one place.

Closes #76.

Why

The dataset adapters (DiDeMo, HiREST) score a single modality against an official evaluator through the benchmark-ready core. Nothing evaluated the complete public product path — fusion, ranking across modalities, and evidence delivery — so a regression there was invisible to any reproducible measurement. Per the issue, existing component benchmarks are reused but do not substitute for evaluating the whole path.

What it adds

vidxp.benchmarks.end_to_end — a transport-neutral harness that drives the same public search operation a real client calls and reports the five measures the issue asks for:

Requested measure	Metric
Relevant moments found	recall_at_k at a relevance IoU (default 0.5)
Timestamp / temporal-range accuracy	mean_top1_iou, per-case best_iou
Ranking quality across modalities	mean_reciprocal_rank, IoU-graded ndcg_at_k, modality_contribution
Evidence supports the result	evidence_support_rate over delivered evidence ranges
Latency, failures, degraded/partial	latency_ms_*, failed_cases, no_result_cases, per-case evidence_degraded

Design. The evaluator is dependency-injected: a caller supplies a SearchFn mapping an EvaluationCase to a public FusedSearchResult. application_search_fn binds the real VidXPApplication.search into that shape, building the same SearchCommand a client sends — so the harness measures the public product path, not an internal shortcut. Because the core is injected, every metric is unit-tested with a deterministic fake and no models. A case whose search fails is recorded as failed and excluded from the quality metrics, so one broken query never hides the rest (application_boundary only wraps known errors, so the harness catches any exception to stay robust).

Scores keep VidXP's ordering-only meaning; calibrated scoring is a separate concern (#90).

Scope

This PR lands the tested core, the public-path adapter, and docs. Wiring the harness to a CLI / HTTP / MCP surface is intentionally left as a follow-up so the tested core lands first — application_search_fn already shows the exact binding a surface adapter needs. Happy to follow up with the CLI command if you'd prefer it in-scope.

Testing

• ruff check src tests — clean.
• python -m unittest discover -s tests — 667 passed, 7 skipped (env-gated ffmpeg/Pillow/symlink), 0 failed.
• 12 new tests cover the metric functions (temporal IoU, IoU-graded nDCG, media-scoped matching), aggregate scoring, evidence support/degradation, failure isolation, no-result handling, parameter validation, and the public-command adapter. New tests are unittest.TestCase subclasses so CI's unittest discover collects them.