Remove Ollama/LLM text cleanup, keep deterministic cleanup

[arankine9]

What & why

Wave's cleanup was actually two stacked layers:

1. DeterministicCleanup — a pure-Swift pass (disfluency removal, spoken-symbol substitution for code, identifier reassembly, spacing/casing) that runs on every dictation with no network.
2. Ollama/LLM polish — an optional stage layered on top that required a local Ollama install.

Since Ollama was never set up, every dictation already ran on the deterministic pass alone. This PR rips out the entire LLM/Ollama layer and all of its UI surfaces, while keeping the deterministic pass running silently — so pasted output is unchanged.

Removed

• Core: OllamaClient, OllamaHealth/OllamaHealthProbe, CleanupClient, CleanupPipeline, CleanupResult, CleanupError, IdentityCache, SystemPrompt.
• Prefs/env: CleanupMode, cleanupModel, ollamaURL, and the WAVE_CLEANUP_MODEL / WAVE_OLLAMA_URL / WAVE_CLEANUP_MODE env vars.
• Dead gate: SkipGate.shouldSkipCleanup (the LLM-skip gate) and AppPaths.identityCacheURL. SkipGate.looksLikeCodeDictation is kept — it still routes prose vs. code in DeterministicCleanup.
• UI: Settings → Cleanup section (mode picker, Ollama model/URL fields, health probe); History panel "Show original" before/after toggle and the Cleaned/Raw badge; onboarding "cleaned text" page reworded to plain paste.
• Tooling/tests: scripts/judge.sh, the cleanup-pairs / identifier-spelling / hallucination-audit fixtures, and the LLM-only tests (CleanupPipelineTests, CleanupModeTests, OllamaHealthProbeTests, SystemPromptTests, TokenBudgetBenchmarkTests, HallucinationFixtureTests, IdentityCacheTests, SkipGateTests, FixtureGateTests).

Changed

• DictationOrchestrator calls DeterministicCleanup.transform directly (the .cleaning status and timing are preserved).
• DictationTrace / HistoryLine drop path / inputTokens / outputTokens. Existing history.jsonl still decodes — JSONDecoder ignores the extra keys.
• Remaining tests rewritten to drive deterministic cleanup with no stub LLM client.
• README / CHANGELOG / SHIPPING / TODO and the install/bench scripts updated to drop Ollama setup and the LLM gates.

Notes / judgment calls

• Output is unchanged for plain prose and spoken code — the deterministic pass was already doing all the work.
• Onboarding step kept, not deleted: I reframed the paste step (removed the "Ollama cleans your transcript" copy) rather than removing the whole page, since it still usefully teaches the core "your words land in the focused app" behavior. Happy to drop the step entirely if you'd prefer.
• rawTranscript retained in the history log (now hidden in the UI) — it's invisible to users but powers the HistoryGoldenTests replay tool.
• Historical entries in CHANGELOG.md / TODO.md describing the old LLM build are left as a record, with a note pointing at the removal.

⚠️ I couldn't compile here (Linux container; this is a macOS/AppKit project). Edits were made against a full read of the code and verified with a repo-wide symbol sweep, but CI (swift test on macOS) is the real check.

---

Generated by Claude Code