bench: categorical failure-mode metrics + Mandarin corpus expansion

[larryxiao]

SPEC: SPEC-021 (Mandarin auto-detect failure: benchmark coverage + fix)

Milestone: M1

Why

Issue #17 shows Mandarin audio transcribed as [SPEAKING CHINESE]. WER/CER alone
can't distinguish three distinct failure modes (placeholder hallucination, silent
translation, garbled output) — they all show elevated numbers but require different
fixes. This PR adds the metrics we need to measure each mode independently and
establishes a baseline before any fix lands.

Change

Sources/OpenQuackBench/BenchRunner.swift

• New FailureMode enum: .placeholder, .silentTranslation, .garbled, .ok
• ClipMetrics gains three fields: failureMode, outputScriptMatch, hallucRateRaw
• BenchResult.failureCounts computed property for aggregate reporting
• classifyFailureMode() — priority-ordered classifier:
  1. Bracket annotation regex ([SPEAKING CHINESE], [FOREIGN LANGUAGE], etc.)
  2. CJK reference + Latin-dominant output → silentTranslation
  3. WER > 200% → garbled
• outputScriptMatch() — true when reference and output share the dominant script
• hallucRateRaw() — fraction of output chars inside [...] brackets
• Unicode block helpers for CJK (Unified + Ext-A + Compat, Hangul, Hiragana, Katakana) and Latin

Sources/OpenQuackBench/Report.swift

• Aggregate table: new Failures column showing PH:N ST:N G:N or all ok
• Per-clip detail: new Mode column (✓ for ok, bold name for failures)
• CSV: three new fields failure_mode,output_script_match,halluc_rate_raw

bench/corpus/fetch.sh + bench/corpus/multilingual/zh_00{3-8}.txt

• Mandarin corpus: 2 → 8 clips via say -v Tingting, covering conversational
  speech, numerals, proper nouns, and instruction sentences

Tests

• swift build && swift test green (CI: macOS-15 + Xcode, where #Preview macros resolve)
• bash bench/corpus/fetch.sh generates zh_003–zh_008 WAVs cleanly ✓
• Bench delta: run before landing PR-B to capture baseline failure rates

Out of scope

• The fix (token suppression, script-match retry) — PR-B, depends on this baseline
• Japanese, Korean, and other-script failure tracking — follow-up once zh fix validated
• Full multilingual Settings UI — M2/M2.5

Closes part of issue #17 (baseline measurement); fix lands in PR-B.