fix(nudge): fire tier-2/3 distillation nudges on block COUNT, not summary token mass

[ranxianglei]

Problem (billion-context-pi#224)

tier2Trigger / tier3Trigger (defaults 5 / 10) existed in config and validation, but decideNudge never read them. The non-pressure path gated T2 on:

t2Pen >= tier2Threshold /* 1.5x nudgeGrowthTokens */ && t2Pen > t1Eff

Summaries are ~10:1 condensed by design, so t2Pen almost never reaches the 1.5x gate, and any fresh raw content keeps t1Eff larger than the summary total. Net effect: the T2 nudge effectively never fires and tier-1 blocks accumulate indefinitely — the user-visible "t2 永不触发" in billion-context-pi#224.

Empirical repro on 0.0.45 defaults (1M limit): 10 tier-1 blocks with ~1K-token summaries + raw messages present → no nudge (max compressible 17136 < threshold 50000); 80 blocks → T1 wins; only the corner case (all raw already compressed) fired T2.

Fix

• Count-based trigger: active tier-1 blocks >= tier2Trigger (T3: tier-2 blocks >= tier3Trigger) now ORs with the existing token-mass path.
• T1 raw compression keeps priority when its effective pending reaches nudgeGrowthTokens (fresh content still gets compressed first).
• Per-tier growthFloor cadence still throttles re-firing.
• Diagnostics: count-ready tiers are listed as T2 5 blocks (count); cadence-blocked tiers show blocked: T2 (cadence).
• TierConfig docstrings document the count semantics.

Verification

• 5 new tests in tests/nudge.test.ts: count-triggered T2 (fires with 5 blocks despite ~500-token summary mass), negative below trigger, T1-priority preserved, cadence respected, count-triggered T3.
• Full suite: 539/539 pass; tsc --noEmit clean; build clean.
• Repro re-run against this build: scenario A (10 blocks + raw present) now fires T2 distill ready: 10 tier-1 blocks >= tier2Trigger 5; scenario B (large raw pending) still routes to T1.

[github-actions]

📦 Built Package Artifact

Branch: 2026-08-25_tier-count-trigger (58863c7)

Option A — Install from npm PR tag (recommended)

In your adapter project:

npm install acp-kernel@pr-162

Each push to this PR publishes a new version under the pr-162 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf acp-kernel-pr162.tgz
npm install ./package

---

This comment is automatically updated on each push.

[ranxianglei]

Agent reviews (2 independent reviewers)

Both verdicts: approve-with-nits. Follow-ups applied in 58863c7.

Correctness review — verified

• Tiers-disabled safety: count conditions inside config.tiers.enabled && (...) (src/compress.ts:1165, 1180) — cannot fire with tiers off. Now pinned by a new test.
• T1 priority intact: t1Eff >= nudgeGrowthTokens checked first (src/compress.ts:1161); count-T2 lives in the else if.
• Pressure/overLimit path precedes the growth chain and ignores count triggers (src/compress.ts:1132–1159) — no interference.
• No empty-target crash: validateConfig enforces triggers >= 1 (src/config.ts:79–81); nudge-text falls back ?? "b1".
• No spurious fires: requires growthReady; first turn growth = 0. Cadence stamped on inject (src/compress.ts:607–611), checked at 1169–1172.
• Re-baseline scale drop clears lastShownByTier (src/compress.ts:594) — count trigger re-arms cleanly.
• Counts derive from activeBlocks(state) computed fresh per decideNudge — syncBlocks deactivation reflected same-turn.

Design review — highlights

• Count is the right signal: README always documented count semantics; token-mass gating was a dimensional error (comparing ~10:1-condensed summaries to a raw threshold). Count also proxies structural overhead (search docs, block metadata) not captured by summary tokens.
• No T1 starvation: worst case a one-turn T1 delay; throttled by growthReady + per-tier cadence (growthFloor ≥ max(20000, 0.45× nudgeGrowthTokens)); compliance resets count (5 T1 blocks → 1 T2). No spam loop found.
• Compatibility: validateConfig unchanged; NudgeState shape unchanged; old semantics were dead config so no user config changes meaning. Release-note that T2/T3 nudges now actually fire.
• Restore-vs-revert: keeps the 2e9efdf unified model (cadence, suppression, pressure routing) and ORs in count — hybrid, not a revert.

Nits addressed in 58863c7

• Dedented the diagnostics else-branch (was 6-space vs surrounding 5-space).
• Added tiers disabled -> count trigger cannot fire T2 test.

Noted, not blocking

• Pre-existing: T3 token path reuses tier2Threshold (no tier3GrowthMultiplier exists).
• Optional future: minimum-mass floor alongside count to skip negligible distillations (low priority given cadence throttle).
• A tier can appear in both ready (count) and blocked (cadence) hints simultaneously — factual but slightly noisy.

[ranxianglei]

📦 Built Package Artifact

Branch: 2026-08-25_tier-count-trigger (58863c7)

Option A — Install from npm PR tag (recommended)

In your adapter project:

npm install acp-kernel@pr-162

Each push to this PR publishes a new version under the pr-162 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf acp-kernel-pr162.tgz
npm install ./package

---

This comment is automatically updated on each push.

[ranxianglei]

[system] ❌ ework-daemon crashed (exit 143). Try posting again or @ework-daemon to retry.