feat: improve Memory v5 task/chat classification with LLM classifier

[XTLine]

Closes #5486

Summary

Improve Memory v5 task/chat detection logic, addressing false positives and false negatives in the keyword-matching approach.

Implemented in two phases:

Phase 1: Keyword-based Gating (first 3 commits)

1. 59dc039 - Throttle memory compiler prompt injection

   • Add gating logic in memorycompiler/runtime.go
   • Use shouldInjectPrompt() to decide whether to inject the contract
   • Add E2E tests to verify throttling behavior

2. bc9b7f2 - Hide memory compiler citations in transcript

   • Filter Memory v5 citation display in frontend Message.tsx
   • Add memoryCitationVisibility.ts utility function
   • Add unit tests for citation visibility logic

3. 76852fe - Gate memory compiler prompt injection for chat

   • Add shouldInjectMemoryCompilerContractForInput() in agent.go
   • Use keyword whitelist (greetings) and blacklist (action words)
   • Support EN/ZH detection: hello, hi, 你好, fix, 修复, etc.
   • Add unit tests in memory_compiler_gate_test.go

Problem: The keyword approach suffers from false positives ("thanks for fixing" → task) and false negatives ("auth isn't working" → chat).

Phase 2: LLM-based Classifier (last 3 commits)

4. d916d0c - Add task/chat classifier infrastructure

   • New task_classifier.go:
     • TaskClassifier interface
     • llmClassifier: classifies via Haiku model (2s timeout)
     • heuristicClassifier: improved keyword fallback
     • classificationCache: session-scoped cache (LRU + 5min TTL)
   • Add full unit tests (including mock provider)

5. bbc2d87 - Integrate classifier to gate Memory v5

   • Integrate classifier in agent.go:
     • Chat inputs completely skip Memory v5 (not just suppress injection)
     • Fall back to heuristic classifier on LLM error
   • Add feature flag in boot.go:
     • Env var: REASONIX_MEMORY_COMPILER_LLM_CLASSIFICATION
     • Default: false (gradual rollout)
   • Clear classifier cache on SetSession()

6. 097a49d - Update and add E2E tests for classifier behavior

   • Update TestRunDoesNotInjectMemoryCompilerContractForGreeting:
     • Expect 0 stats events (chat completely skips Memory v5)
   • New tests:
     • TestClassifierSkipsMemoryV5ForChat
     • TestClassifierUsesMemoryV5ForTask

Cache-impact: low - New session-scoped cache (max 100 entries, ~10KB) only active when feature flag enabled (default: false). Cleared on session boundary.

Cache-guard: Comprehensive unit tests in internal/agent/task_classifier_cache_test.go covering Get/Set, TTL expiration, LRU eviction, normalization, and Clear behavior. Run: go test ./internal/agent/... -run TestClassificationCache -v

System-prompt-review: No system prompt content changes. Modified boot.go only adds feature flag configuration
(REASONIX_MEMORY_COMPILER_LLM_CLASSIFICATION) without altering prompt templates or agent instructions.

Verification

All tests pass:

  go test ./internal/agent/... -v

Test coverage

• Heuristic classifier unit tests (greetings, tasks, false positive/negative scenarios)
• LLM classifier unit tests (mock provider)
• Cache behavior tests (Get/Set, TTL, LRU, normalization)
• E2E tests (chat skips, task uses Memory v5)
• Gating tests (keyword-matching logic)

Cache details

New cache: classificationCache (only when feature flag is enabled)

• Scope: session-scoped (cleared on SetSession())
• Size: max 100 entries (LRU eviction)
• TTL: 5 minutes
• Key: SHA256(normalized_input)
• Memory footprint: ~10KB (100 entries × ~100 bytes/entry)

Cache strategy:

• Hit: return directly (<1ms)
• Miss: call LLM (~100-300ms) → cache result
• LLM error: fall back to heuristic → do not cache error results

[SivanCola]

Approved. This is now a well-scoped fix for the Memory v5 chat/task noise reported in #5486 and the related user-visible compiler citation noise.

The selected mechanism is the right one: Memory v5 startup is gated by task/chat classification, the heuristic classifier remains the default path, the optional LLM classifier is explicitly opt-in, simple greetings and acknowledgements skip execution-contract injection, real task/failure descriptions still use Memory v5, and Memory v5 compiler citations are kept out of the visible transcript.

Why this fits the architecture:

• The default path stays local and cache-friendly; it does not add provider calls unless the opt-in environment flag is enabled.
• The classifier controls the Memory v5 startup/injection boundary instead of only decorating it, so non-imperative task reports like "the auth isn't working" still get the compiler path.
• It remains compatible with #5487: #5487 filters backend evidence citations, while this PR handles chat/task gating and transcript visibility.

The main thing I wanted tightened was the classifier-to-injection contract and the boundary coverage. The follow-up commits now cover the Memory v5 gating path, classifier caching, heuristic chat exclusions, Chinese task phrases, failure semantics, English boundary cases, and frontend citation visibility. Related older work in #5376 is superseded by this PR's newer implementation path.

Verified:

• latest GitHub checks are green: CI, race, desktop, lint, coverage, govulncheck, CodeQL, and cache-impact
• clean local merge against current main-v2 succeeded
• go test ./internal/agent/... ./internal/memorycompiler/... passed after that merge
• git diff --check passed