test(worker): V2.1 contract suite + pipeline extraction + 2 fixes

[k8-benetis]

What

• Extracts the pure processing pipeline (gaps/LTTB/extrema/domains/evictions) to src/workers/pipeline.ts — mechanical move, verified verbatim; the worker keeps cache/fetch/orchestration/onmessage.
• Adds 43 tests (65 total): math ground truths + full protocol harness (fake self, stubbed fetch, per-test module isolation) pinning the Contract V2.1 invariants — monotonic xs, NaN only at gap bridges, per-series payloads, complete transfer list, cache-not-detached, retry-on-empty, mixed sec/ms/us epochs, error taxonomy.
• Fix 1 (confirmed bug): the cache stored POST-processed buffers keyed without the processing policy — a viewport resize (new downsampleThreshold) served data processed for the old viewport. The key now appends |g{maxGap}|t{threshold}|e{extrema} (unused viewportWidthPx/compatibilityMode excluded; key stays opaque to the UI).
• Fix 2 (confirmed bug): gap-bridge NaN survival under downsampling was bucket-alignment luck — for maxGap < gap < 2×maxGap the bridge sits inside an LTTB segment and was frequently dropped, so the rendered line spanned real sensor outages (violating acceptance criterion 3). downsampleSingle now force-retains NaN indices (O(gaps), negligible).

Notes

• Worker is inline-bundled (?worker&inline) — no rolling-deploy version skew.
• Contract doc updated to V2.2 in internal docs (V2.1 marked Superseded), including post-spec behavior now contractual: URN passthrough (Strangler), retry-on-empty, mixed-epoch normalization, mean downsampleRatio aggregation, concurrency semantics.
• Pre-existing findings out of scope (follow-ups): PanelSeriesRail per-series stats lookup compares a 3-part key against the worker's full cache key (dead since V2.1); useWorkerSeries.release() has no call sites.

Suite: 65 passed locally, deterministic (two runs); typecheck clean.