Summary

Revises the training trigger and sampling model for S1.2 implementation, based on design discussion with @ikolomi.

Changes

Removed:

• compression-dict-first-training-keys-count — replaced by compression-dict-min-training-keys

New/updated knobs (all configurable):

Cooldown:

• 30s hardcoded cooldown after a failed training attempt (scan didn't collect enough eligible samples). Avoids expensive repeated scans when the DB has many non-eligible keys (INT, EMBSTR, too small, etc.).

Flow

compressionCron:
  total DB keys >= min (1000)?
  AND no training in progress?
  AND not in cooldown?
  → start scan

Scan collects eligible values (raw strings within size bounds).
Stops when:
  samples == max (10K)  OR  buffer == 16 MiB  OR  keyspace exhausted

After scan:
  eligible samples >= min (1000)?  → submit to bio for training
  eligible samples < min?          → abort, enter 30s cooldown

Why

• Trigger uses total DB keys (not eligible keys) because eligibility requires scanning — we can't know without doing the work. Total key count is O(1) via kvstoreSize.
• Buffer cap (16 MiB) prevents memory explosion from large values.
• Key cap (10K) bounds scan duration; beyond 10K ZSTD quality saturates.
• Min (1000) ensures enough diversity for a useful dictionary.
• Cooldown (30s) prevents wasting main-thread CPU on repeated failed scans.

Design doc changes

• R2.3.5: trigger uses compression-dict-min-training-keys, 30s cooldown on failure
• R2.3.6: dual-cap model (max keys OR buffer size), min floor for submission
• Config table: removed old knob, added 2 new knobs with detailed descriptions, updated count to 12
• §6.3: error handling updated to reference new knob + cooldown

No code changes — design doc only.