Lean-output nudge now fires for everyone
The gentle-tier nudge required both fill >= 20% and quality score < 75. Quality only falls below 75 after a session has already degraded, so the nudge was effectively unreachable outside very long sessions — most users never saw it at any fill level.
Fill is now the only condition, at 25%. A healthy session that is a quarter full still pays for verbose output, so it now gets the nudge. The 3-per-session cap and 5-minute cooldown remain the guards against noise. The 75%+ strong tier and 90%+ suppression are unchanged.
TOKEN_OPTIMIZER_VERBOSITY_MIN_FILLdefault:20→25- Verified end-to-end through the real
UserPromptSubmitentrypoint: fill 30 / quality 100 fires, fill 25 fires, fill 24 silent - Full suite: 374 passed. OpenCode port: 74 passed.
Platform parity: Claude Code, Codex, Hermes and Copilot share measure.py, so the fix propagates to all four. The OpenCode port (token-optimizer-opencode 1.1.1) was the only second implementation and is now identical. OpenClaw's copy of this feature is still status: deferred; only its registry description was corrected.
Docs realigned across README and three docs-site pages.