fix(memory-core): make remote embeddings token-aware and resilient

[gavinlouuu-kpt]

Summary

• add token-aware remote embedding budgets with Unicode-safe truncation
• retry only transient 429/5xx/network/timeout failures with bounded exponential backoff
• validate vector count, dimension, finiteness, and magnitude before persistence
• preserve per-item progress during deferred L0 embedding and retain failed source rows for replay
• expose safe embedding failure counters plus separate L0/L1 source/vector coverage in /health

Verification

• npm test: 5/5 passed
  • CJK + emoji + code against a smaller token budget
  • bounded 500/429 retries
  • bounded timeout retry
  • permanent 4xx no-retry and partial item progress
  • L0 durability across restart and idempotent vector backfill
• npm run build:plugin
• git diff --check

Tracks downstream CDB issue gavinlouuu-kpt/codex-discord-bridge#247.

[gavinlouuu-kpt]

Live downstream staging validation passed with BGE-M3 at 1024 dimensions: eight isolated L0 writes embedded successfully, the extraction/recall pipeline completed, health stayed ready, and cleanup returned source/vector coverage to 1. Evidence SHA-256 f4d2e1de8fe430e3ec10fac87980abd0469339979fc5fa7c19bb37a0577f1215.

[Maxwell-Code07]

Thank you so much for your attention and contribution! We will arrange an internal review for this PR shortly, and all feedback will be shared right here in the discussion.