feat: add ModeClassifier hook — deterministic mode routing

[rikitikitavi2012-debug]

Summary

• Adds ModeClassifier.hook.ts — a deterministic UserPromptSubmit hook that classifies every prompt into ALGORITHM or MINIMAL mode using regex patterns (< 20ms, zero API calls)
• Fixes mode classification regression where ALGORITHM mode activates only ~9% of the time due to CLAUDE.md template attraction bias (the LLM pattern-matches to NATIVE format over semantic instructions)
• Registers the hook as first in UserPromptSubmit chain in settings.json (before RatingCapture)

Problem

The current mode classification relies on the LLM reading CLAUDE.md instructions and choosing the correct mode. In practice, the LLM latches onto the NATIVE format template (which appears first and is simpler) and uses it ~91% of the time, even for complex tasks that should trigger ALGORITHM mode.

Solution

Remove the LLM from mode classification entirely. A deterministic regex hook runs before the AI sees the prompt:

• 39 regex patterns match greetings, ratings, thanks, acknowledgments (EN + RU/Cyrillic) → routes to MINIMAL
• Everything else → routes to ALGORITHM (the default for actual work)
• The Algorithm's Complexity Gate in OBSERVE phase then handles intelligent downshifting to NATIVE for genuinely simple tasks

Design Decisions

• Regex over LLM: Deterministic, < 20ms, zero cost, zero hallucination
• Conservative MINIMAL matching: Only unambiguous patterns (anchored ^...$). "ok" matches but "ok, do X" doesn't
• Position 0 in hook chain: Must run before other UserPromptSubmit hooks for correct mode injection
• Fail-open: Any error → defaults to ALGORITHM (safe fallback)

Test plan

• Verify hook runs on UserPromptSubmit event
• Test English greetings → MINIMAL (hi, hello, hey, thanks, good morning)
• Test Russian greetings → MINIMAL (привет, здравствуй, спасибо, доброе утро)
• Test ratings → MINIMAL (9/10, 8.5, "7/10 норм")
• Test complex prompts → ALGORITHM (any actual work request)
• Test edge cases: "ok" → MINIMAL, "ok, deploy to prod" → ALGORITHM
• Verify < 20ms execution time (no API calls, no file I/O)
• Verify fail-open behavior on malformed input

🤖 Generated with Claude Code

[rikitikitavi2012-debug]

Friendly ping — this PR adds a ModeClassifier hook that handles deterministic mode routing (MINIMAL/NATIVE/ALGORITHM) via regex before LLM evaluation, saving tokens on simple inputs like greetings and ratings.

Happy to address any feedback or adjust the approach. Let me know if there's anything blocking review.

[rikitikitavi2012-debug]

Friendly ping — this PR adds a ModeClassifier hook that routes requests to the correct mode (Algorithm/Native/Minimal) via fast regex before LLM processing, reducing unnecessary LLM calls. Mergeable, no conflicts. Happy to adjust if needed.

[rikitikitavi2012-debug]

Friendly ping — this PR has been open for a few days without review. Happy to address any feedback or make changes if needed. Let me know if there's anything blocking the merge.