[ranxianglei]

Fixes for review findings on 0.0.20 (MAJOR-1, MAJOR-2, MINOR-1, MINOR-2)

Oracle review of 0.0.20 found 4 issues in the decideNudge over-limit force path. This PR fixes all 4 + adds test coverage.

MAJOR-1: perpetual nudge spam when nothing is compressible

compress.ts:916 — shouldInject = overLimit || injectedTier !== null fired even when all content was protected/already compressed (injectedTier=null but overLimit=true).

Fix: gate on actionable target:

const shouldInject = injectedTier !== null
  || (overLimit && (rec?.recommendedRanges?.length ?? 0) > 0);

Now shouldInject is false when there's truly nothing to compress — the truncate node (at truncate.threshold) is the safety valve for that case, not a nudge.

MAJOR-2: over-limit nudge rendered as gentle

nudge-text.ts:149 — voice decided by emergencyOverride only; the new [0.75, 0.95) force-nudge band rendered gentle with header "not an overflow warning" — self-contradicting at 94% usage.

Fix: isEmergency = !!breakdown?.emergencyOverride || !!breakdown?.overLimit — over-limit now uses emergency voice.

MINOR-1: force loop accepted sub-minCompressRange ranges

compress.ts:907 — info.pending < 1 accepted any non-zero tier, including ranges below minCompressRange that applyCompression would reject.

Fix: info.pending < config.compress.minCompressRange.

MINOR-2: misleading comment

Comment claimed emergencyThresholdPct "trips the truncate node" — they're independent knobs.

Fix: rewritten to state independence.

Tests added (3 new, 230 total)

• nudge.test.ts: "over-limit fires force-nudge when compressible content exists" + "over-limit does NOT inject when nothing is compressible (MAJOR-1 fix)"
• nudge-text.test.ts: "over-limit renders with emergency voice (MAJOR-2 fix)"

Pre-flight

typecheck clean / 230 tests pass / build success