fix(cache): compact newest tool results first to preserve prompt cache prefix

[bcherny]

Problem

compactExistingToolResultsInPlace (src/agents/pi-embedded-runner/tool-result-context-guard.ts:111) iterated front-to-back. When context exceeded 75%, it replaced the oldest tool results with [compacted: ...] placeholders first. This rewrote messages[k] for small k, invalidating the provider prompt cache from that point onward on every turn past the threshold.

Fix

Reverse the loop to compact newest-first: for (let i = messages.length - 1; i >= 0; i--). The cached prefix stays byte-identical; only the tail diverges.

Tradeoff

The model loses recent tool output instead of old. Acceptable because:

• This guard only fires as an emergency measure past 75% context
• Cache-bust cost (full re-tokenization every turn) outweighs the model-quality hit
• If compaction isn't enough, the 90% preemptive-overflow cascade still triggers full session compaction

Verification

• 11/11 tests pass in tool-result-context-guard.test.ts
• No downstream dependency on compaction order: sole caller enforceToolResultContextBudgetInPlace only reads the return value (a sum) and the in-place mutations
• Early-exit (reduced >= charsNeeded) is direction-agnostic — the accumulator is a simple sum
• applyMessageMutationInPlace + cache invalidation are per-message, no ordering assumption

Notes

• Test assertions updated to reflect newest-first ordering (the "survives repeated large tool results" test previously asserted oldest compacted / newest intact; now inverted)
• Pre-commit hook currently fails on origin/main due to unrelated getEditorKeybindings import breakage introduced in c6d8318 — this change does not touch those files

🤖 Generated with Claude Code