Ponytail minimization pass: dead code, formatter locale fixes, redundant-work cuts

[r3dbars]

Why

Repo-wide minimalism pass (the ponytail decision ladder: delete > reuse > stdlib > native > shrink) targeting performance and reliability. Every finding was adversarially verified against the repo's landmines (source-text-pinned tests, CoreAudio real-time rules, failed-audio retention, privacy sanitizers) before being applied; refuted findings were dropped.

Product Impact

• Affects: dictation / meetings / agent artifacts
• Lane: dictation reliability / meeting reliability
• Why this matters: removes ~1,430 net lines of dead and duplicated code, fixes two real formatter-locale correctness bugs on non-Gregorian system calendars (transcript frontmatter feeds audio-retention decisions; stats-DB date range queries), and cuts measurable redundant work (repeated file reads/parses, per-row SQL prepares, per-call regex/formatter allocation, per-render filesystem stats).

What changed

• Deleted the legacy SystemAudioCapture CoreAudio process-tap path (791 lines, zero production callers — SCKAudioCapture is the only backend) plus its orphaned CoreAudioUtils helpers; docs updated
• Locale-pinned frontmatter/stats dates: transcript date:/time: and six stats-DB sites now write via the POSIX-pinned DateFormattingHelper, matching their POSIX-pinned readers
• Meeting audio-retention maintenance parses each transcript's frontmatter once per pass instead of 2–3×
• FailedTranscription: audio/error fields are var, collapsing five 10-field copy-with-mutation sites that could silently drop a future field; same with(...) shape for RecordingHealthInfo
• Observability: one shared crash-safe log-prepare sequence for EventReporter + ReliabilityPacketRecorder, one shared duration_bucket helper; the pinned source-text test now guards the shared helper
• ContextCaptureEngine: the mirrored _sharedSessionController global became an explicit isHotkeyRoutingActive flag — same late-CGEventTap-event protection, no triple manual sync
• Dead code: old "Call Recording" saver + formatMarkdown, SpeakerProfileMerger name-merge API + variants table (MCP keeps its standalone mirror), unused wrappers in TranscriptionTaskManager / RetroactiveSpeakerUpdater (moved to a test helper) / ParakeetEngine / ParakeetModelLifecycle / WhisperEngine, and an NSLock.withLock shadow of Foundation's own
• Hot loops: regex/DateFormatter hoists (DictationTranscriptStore, RecentCaptureScanners, CaptureMarkdownParser, HomeView), single-evaluation Speakers settings render, lazy Agent-page disclosure content
• Tools: CLI reads each meeting file once (was 3×); MCP receipts tools stop double-hydrating titles (2× disk reads), dictation reads read once, SQLite indexer prepares INSERTs once per transaction; QA dedups process termination, report writing, and the isolated-launch env block

How I checked it

• scripts/dev/agent-preflight.sh
• Selected checks from .agents/test-matrix.yml for the files changed
• bash build.sh --no-open (launch-to-interactive 906ms / 3000ms budget)
• bash run-tests.sh (12,203 passed)
• Performance budget passed (bash build.sh --no-open runs the bundle gate)
• bash run-integration-smoke.sh if I touched Sources/Meeting/ or Sources/TranscriptedCore/
• swift test if I touched Package.swift, Sources/TranscriptedCore/, or the public core seam (1,036 passed, 13 hardware-skips)
• Manual check: bash run-e2e-smoke.sh + all four tool package suites (CaptureKit 53, CLI 78, MCP 192, QA 65) + scripts/dev/check-build-source-lists.py

Risk Review

• Privacy / local-first behavior reviewed (sanitizers/allowlists untouched; shared log-prepare keeps restrict-before-open, pinned test updated to guard it)
• Storage path or migration impact reviewed (no path or schema changes; MCP index schema untouched)
• Public-facing copy stays concrete and matches current product scope (no copy changes)
• Release/update impact reviewed (no release-path or version changes)
• Agent PRs link the issue/workpad and stay draft until human review (landing directly per owner request in-session)
• UI changes include sanitized .agent-review/visuals/ evidence (UI changes are behavior-preserving refactors: identical view output)
• No private transcripts, audio, tokens, personal paths, or customer data are included

Notes

Deliberately skipped (verified-risky or refuted): caching a meeting title column in the MCP index (real perf win, needs a schema migration — good follow-up), two large ParakeetEngine extractions (the claimed-identical paths differ by an asrManager guard), CameraActivityMonitor/MicActivityMonitor scheduler unification (documented intentional parallelism), TranscriptedStoragePaths caching (freshness checks are load-bearing), and CaptureLibrarySize/ModelCacheInventory unification (symlink semantics differ).

Environment note: this machine needed Xcode's Metal Toolchain component (xcodebuild -downloadComponent MetalToolchain, 687.9 MB) for build-deps.sh to compile the MLX shaders — installed during verification. Also, build-deps.sh/build.sh failures are easy to mask by piping through tail (staleness message prints even on the success path of a later run); worth a follow-up to make the stale-deps case louder.

Agent handoff

COORD_DONE: GREEN | (this PR) | ponytail minimization across Sources+Tools, −1,430 net lines, 2 locale-correctness fixes | none | none | build-deps, build, run-tests (12,203), integration smoke, swift test (1,036), e2e smoke, 4 tool package suites, source-list check | merge

🤖 Generated with Claude Code