[ranxianglei]

Problem

Diagnosed from the ACP endurance report (dog/billion-context-pi#23): under a
24.5h isolated session, compress failure rate hit 60.5% (825/1363), and the
dominant cause was "too small" rejections — 707 cases, 86% of all failures,
476 of those single-range.

Root cause is a granularity mismatch between nudge arbitration and the
ranges the nudge offers the model:

• pendingByTier computed T1 pending as the aggregate sum of all
  compressible ranges (inflated by fragmentation).
• That aggregate let T1 win the arbitration loop, so the nudge fired at T1.
• But the nudge then offered per-turn ranges (each often < 5000 chars after
  buildCompressibleRanges splits at user boundaries), and the model picked
  only 1–2 of them → `minCompressRange` rejection.
• Emergency was worse: it skipped the tier loop entirely and always pointed at
  T1-only ranges, never at tier distillation — so T1 blocks piled up unconsolidated.

Fix

1. `mergeRangesToThreshold` (recommend.ts) — two-pass grouping. A greedy
accumulator merges consecutive compressible ranges until the batch reaches
`tokens*4 >= compress.minCompressRange`, then closes. Merges across ref gaps;
count-weighted `toolPct`; `dangerous` propagates. `recommendedRanges` now
returns these merged ranges, so the nudge offers batch-sized ranges the model
can compress in one call without tripping `minCompressRange`.

2. Effective T1 pending (compress.ts) — `pendingByTier` now counts only
effective merged ranges (`tokens*4 >= minCompressRange`) for T1, killing
the inflation that starved tier distillation.

3. Tier-aware arbitration (compress.ts `decideNudge`) —

• Non-emergency: T1 is default when `t1Eff >= nudgeGrowthTokens`; T2
  overrides only when `t2Pen >= nudgeGrowthTokens × 1.5` AND `t2Pen > t1Eff`.
  New config `tier2GrowthMultiplier: 1.5`.
• Emergency: ignores tier priority, picks the tier with the max pending
  (T1 effective vs T2/T3 summary totals) — so emergency now CAN route to
  distillation when that reclaims the most tokens.

4. `renderNudgeText` (nudge-text.ts) — tier≥2 path now also serves the
emergency case: emergency voice + `[EMERGENCY — TIER N DISTILLATION]` header,
still with the full distillation block list + rules.

Tests

+16 (239 total, all pass; typecheck clean; build 103.72 KB).

• `recommend-merge.test.ts` (7): empty / disabled / single≥ / two-merge /
  three-split / dangerous propagation / count-weighted toolPct.
• `nudge.test.ts` (5): non-emergency T1; non-emergency T2 @1.5× && >T1;
  T2 large but T1 wins; emergency T2>T1 → tier 2 + override; emergency T1 max → tier 1.
• `nudge-text.test.ts` (1): emergency + tier 2 → emergency voice + distill text.
• `validation.test.ts` (1 bonus): batch aggregate of small ranges passes when
  total ≥ `minCompressRange` (documents the batching the report questioned).

Compatibility

No adapter API change. `tier2GrowthMultiplier` defaults to 1.5; existing
config-overrides continue to work (it is additive). Once merged and published,
billion-context-pi will bundle it on its next release bump.

[ranxianglei]

Cross-link: complementary with the adapter-side filter

pi-side adapter filtering (ranxianglei/billion-context-pi#148 + the extracted billion-context-kit viableRanges, ≥200 tok) attacks the same fragmentation root from the presentation layer. With this PR merged, that filter degrades gracefully into a safety net (tail leftovers + older-kernel repos, e.g. opencode still pinning 0.0.19). No conflict — merged ranges (≥ minCompressRange ≈ 1250 tok) pass the adapter floor trivially; the kit filter only trims the unmergeable tail groups.

Review: two regressions found

1. The maxContextLimitPct band was deleted → dead config (high).

Master had three arbitration bands: growth-gated normal, overLimit (usage ≥ maxContextLimitPct, default 0.75) bypassing the growth gate with any tier pending >= minCompressRange, and emergency (≥ emergencyThresholdPct, 0.80). This PR keeps only growthReady + emergencyOverride — grep overLimit|maxContextLimitPct in the new compress.ts: 0 hits. Consequences:

• usage in the 75–80% band with growth < floor no longer nudges at all (master fired immediately, bypassing growth);
• maxContextLimitPct survives in config.ts/types.ts with no consumer — dead config inviting misconfiguration;
• the shouldInject = … || (overLimit && recommendedRanges.length > 0) last-resort fallback is gone.

2. Emergency with nothing effective now injects every turn (medium).

Master: emergency + no eligible tier → shouldInject = false (the "(no compressible content)" reason was computed but never sent). New: shouldInject = emergencyOverride || injectedTier !== null → at ≥80% with zero effective pending, an EMERGENCY nudge fires every turn; worse, if recommendedRanges still holds sub-threshold tail groups, the nudge offers ranges that will be atomically rejected — the exact failure mode this PR fixes, resurfacing under pressure. Master's overLimit gate (pending >= minCompressRange) existed precisely to prevent offering sub-threshold ranges.

Suggested fix: restore a mid band — treat usage >= maxContextLimitPct like emergency but gated on effective pending (any tier > 0), and only inject emergency when some tier has effective pending or ranges are nonempty; keep a distinct reason suffix for the no-content case so the 75–80% early pressure isn't silent.

Checked clean

• mergeRangesToThreshold merging across gaps is safe at compress time: soft-protection excludes + warns (compress.ts:583-619), covered messages drop out of directMessageIds (their original blocks keep them), protected tool pairs excluded per the Bug 39 handling. Only cosmetic nit: merged count sums compressible parts but the span visually includes gap messages.
• T1 cadence-check removal is fine — the growth gate (growthReference = lastNudgeShownTokens) is the same condition in disguise.
• tier2GrowthMultiplier is additive; adapters spread-merge user config, no downstream break. NudgeConfig addition is non-breaking until adapters bump their kernel pin.
• Idle reason format changed (ready/blocked hints) — adapters log it verbatim, no parser depends on it; only e2e tests asserting exact idle strings would notice.

239 tests pass because nothing covers the 75–80% band or the emergency-with-nothing case — both worth a test before merge.

[ranxianglei]

Fixes for both regressions are up in #70 (contains this PR's commits via merge + current master, which also resolves the stale emergencyThresholdPct: 0.80 against master's 0.95 truncate tuning). Two new regression tests are red on this PR as-is and green on #70. Suggest closing this in favor of #70 after CI.