NaturalLanguage entity tagging and CI

[zac]

Summary

Stacked on #16 (retargets to main when that merges).

NaturalLanguage entity tagging

• New EntityTagger provider protocol on MemoryConfiguration, with an NLTagger-backed NLEntityTagger as the default under the MemoryNaturalLanguage trait (nil otherwise), mirroring the query-side NLQueryAnalyzer.
• Tagger results merge additively into heuristic extraction: they upgrade generic entity labels to person/location/organization and append entities the heuristics missed in well-capitalized text. Heuristic behavior remains the floor — lowercase chat text leaves the tagger silent, and entityTagger: nil reproduces pre-change behavior exactly. Facet inference now sees merged entities, so location/person facets fire beyond the built-in lexicon.
• Ambiguous self-reported locations ("i'm in lisbon") become durable profile memories only when the tagger confirms the phrase names a place, validated by title-casing it inside a carrier sentence — empirically the high-precision formulation (bare-phrase tagging mislabels; carrier junk like "a meeting"/"fear" is rejected). Travel qualifiers ("i'm in boston this week") mark the mention transient and skip the residence write. Residence statements ("i live in ...") never require tagger confirmation, keeping eval gates OS-stable.
• save/ingest entity normalization consults the tagger when candidates supply no entities.

CI

• ci.yml runs on every PR commit and pushes to main: swift test with default traits plus the leakage check and strict dataset validation, and a separate CoreML+NaturalLanguage trait build job. SwiftPM build dirs cached per trait flavor; in-progress runs cancel on new commits.
• evals.yml is workflow_dispatch-only: extraction-affected suites under coreml_default, pressure/smoke baseline gates, reports uploaded as artifacts. Deliberately not PR-blocking since NaturalLanguage models vary by macOS version.

Testing

• 120 tests pass locally, including new coverage for tagger-confirmed locations, transient-state rejection, label upgrades, lowercase no-op behavior, and the no-tagger heuristic floor.
• All five extraction-affected eval suites reproduce baseline metrics exactly under coreml_default (96.49/85.10/95.29 on schema gold; 100% elsewhere; 0% false-writes) — the default-on tagger causes zero gate drift.
• Leakage checker and CoreML-trait build clean. This PR also exercises the new CI on itself; if runner NaturalLanguage models diverge from dev machines, the NL-dependent tests will surface it here.