feat: hallucination filter — strip known ASR boilerplate before writing (#98)

[kiki830621]

Summary

Adds a backend-agnostic post-decode pass that strips known ASR hallucinations before a transcript is written, so silent-segment boilerplate no longer reaches the output and downstream srt-proofread doesn't have to clean it.

Refs #98.

What it does

HallucinationFilter.filter(_:mode:denylist:) — a pure function over a Transcript, applied at the single CommandCore.transcribe choke point, after diarization:

• Denylist strip — drops cues matching a curated list of known boilerplate. Default set is Whisper-family (the "please like & subscribe" outros that surface verbatim over silence). Matching is normalized (whitespace + punctuation folded), so decoder spacing jitter doesn't need a new entry.
• Empty / adjacent-duplicate collapse — drops empty cues and adjacent exact-duplicate cues (rolling-caption / token echo). Universal across backends.
• Re-indexes surviving cue ids 1…N and re-derives the flat text the same way an engine does (segments.map(\.text).joined()), so srt / txt / json / vtt all come out consistent. Speaker labels on surviving cues are preserved.
• No-op path returns the input untouched (ids + text byte-identical).

Surfaces

• CLI: --hallucination-filter <off|denylist> on transcribe (default denylist).
• MCP: hallucination_filter on the transcribe tool (default denylist).
• GUI: inherits the default via the CommandCore signature (no code change).

Behavior change

Transcription now runs the filter in denylist mode by default — known boilerplate no longer reaches the output. Pass --hallucination-filter off (CLI) / hallucination_filter: "off" (MCP) to restore raw output. Documented in CHANGELOG.

Design notes

• Placed at the single choke point after diarization → covers CLI/MCP/GUI in one place, backend-agnostic, and keeps speaker labels aligned.
• The denylist content is Whisper-family, so it is a harmless no-op for backends that never emit those strings (e.g. Parakeet); the empty/duplicate collapse is universal.
• HallucinationFilterMode lives in BestASRKit (ArgumentParser-free); the CLI adds the ExpressibleByArgument conformance, so --help auto-lists the cases and rejects unknown values.

Out of scope (tracked in #98)

• Phase 2: full / confidence-gated filtering (needs no_speech_prob / compression_ratio plumbed through the segment model, per-backend fallback).
• Phase 3: decode-param knobs (WhisperKit-only).

Tests

HallucinationFilterTests (6 cases): boilerplate drop + reindex, look-alike legitimate speech retained, .off pass-through, empty/adjacent-duplicate collapse, clean no-op, speaker labels preserved. Fixture uses the exact 3-cue evidence string from #98.

Full suite green: 341 tests / 73 suites. swift build clean across all targets (Kit / CLI / MCP / GUI).

Manual smoke test (recommended — needs model download)

Re-transcribe a recording with a known silent-segment hallucination and compare default vs --hallucination-filter off: the boilerplate should be absent by default and present with off.