fix: align nudge summary token counts with tokenizer - #91

Problem

Tier-distillation nudges display each target block's summary token count, but the renderer estimated that value with summary.length / 4.

The decision layer already uses the injected countTokens function for the same summaries. This makes the displayed compression payoff diverge from the value used for tier arbitration, especially for CJK text and custom tokenizers.

Root Cause

pendingByTier had access to the active countTokens, but NudgeDecision only carried the target CompressionBlocks.

The per-block token counts were therefore lost at the decision boundary, forcing renderNudgeText to estimate them again with length / 4.

Fix

• Precompute per-target-block summary token counts in the decision layer using the active injected countTokens.
• Derive pendingT2 / pendingT3 from those same precomputed values so arbitration and rendering share one token-counting source.
• Carry the derived values on NudgeDecision as tierTargetBlockStats.
• Make the renderer consume those values instead of re-tokenizing normal kernel-produced decisions.
• Preserve the previous length / 4 behavior as a compatibility fallback for manually constructed legacy decisions that do not provide stats.

Scope / Non-goals

• CompressionBlock is unchanged; derived token counts are not persisted into runtime state.
• renderNudgeText(decision, prompts?) keeps its existing signature.
• No dependency, version, lockfile, workflow, or documentation changes.
• No unrelated refactors.

Tests

Added regression coverage for:

• CJK summaries using the default CJK-aware tokenizer.
• A custom injected tokenizer that differs from both length / 4 and defaultCountTokens.
• Renderer use of precomputed per-block token counts.
• Legacy decisions without stats.
• ASCII output compatibility.

Validation:

• Targeted tests: 47/47 passed.
• Full suite: 390/390 passed.
• npm run typecheck: passed.
• npm run build: passed.
• git diff --check: passed.

Compatibility

The new NudgeDecision.tierTargetBlockStats field is optional.

Existing manually constructed decisions without the field continue to use the previous renderer estimate, while normal kernel-produced tier-2/tier-3 decisions provide tokenizer-consistent stats.

Limitations

Legacy/manual decisions that omit tierTargetBlockStats intentionally retain the previous fallback estimate for backward compatibility.

Fixes #45