fix(gradient): deterministic timestamps in sanitizeToolParts

[BYK]

Summary

• Replace Date.now() with the tool part's existing start time (or 0 for pending parts) in sanitizeToolParts() to ensure repeated transform() calls produce identical bytes
• Eliminates 196 non-distillation cache busts per long session ($353 savings on analyzed session)

Root Cause

OpenCode's prompt-loop cache fix (e148f00aa) preserves old pending tool parts in the cached message array across prompt loop iterations. Each transform() call, Lore's sanitizeToolParts() re-converts these stale pending parts to error state using Date.now() — producing different serialized bytes on every API call and invalidating Anthropic's prompt cache from that position forward.

All 196 non-distillation busts in session CLI-1AE (1,733 API calls, $1,090 total) followed a tool-calls finish reason and showed cache_read ≈ 41K (only system prompt survived). 154 of 196 were consecutive bust chains.

Fix

Use part.state.time.start for running parts (which already have a timestamp) and 0 for pending parts (which have no time field). Both are deterministic across repeated calls on the same input.

Combined Impact (with PR #123)

On the analyzed CLI-1AE session:

• PR perf(gradient): batch distillation consumption at turn boundaries #123 (consumption freeze): eliminates 230 distillation-correlated busts ($513)
• This PR (deterministic timestamps): eliminates 196 tool-part busts ($353)
• Combined: 426 → ~0 busts, $866 → ~$0 cache-write waste (79% of $1,090 total)

Cache Control Investigation

Also investigated whether explicit Anthropic cache_control breakpoints could help. Findings:

• OpenCode already uses all 4 available breakpoints (2 system + 2 trailing messages)
• The experimental.chat.messages.transform hook operates on MessageV2.WithParts[] before cache annotations are added — cannot inject cache_control from plugins
• The 20-block lookback window works correctly as long as prefix bytes are deterministic, which this fix ensures
• No action needed on cache_control — the existing setup is optimal

Tests

• Added sanitizeToolParts determinism test verifying consecutive transform() calls produce byte-identical output for stale pending tool parts
• Full suite: 686 pass, 0 fail