fix: preflight compact v3 context before model calls

[nguyenha935]

Summary

• Run V3 pruning/compaction before ThinkStage so over-budget histories are compacted before provider calls.
• Persist V3 last_prompt_tokens with the post-flush persisted message count, matching V2 calibration behavior.
• Add regression coverage for preflight compaction and V3 metadata message count.

Tests

• go test ./internal/pipeline -run "TestNewDefaultPipeline_PrunesBeforeThink|TestFinalizeStage_UpdateMetadataReceivesPersistedMessageCount" -count=1
• go test ./internal/agent -run "^$" -count=1

Surface parity

• Gateway runtime: affected, V3 pipeline order and metadata callback updated.
• API contract: N/A, no request/response schema changes.
• Web UI: N/A, no UI behavior or strings changed.
• CLI/runtime package: N/A, no CLI flags or runtime packaging interfaces changed.

[clark-cant]

Clean, focused fix. Pipeline reorder (Prune before Think) is correct — over-budget histories should be compacted before provider calls. UpdateMetadata msgCount addition matches V2 calibration behavior. Good regression coverage.

Verdict: Approve

• Duplicate/prior work: clear — no prior implementation of preflight V3 compaction
• Project standards: compliant
• Strategic necessity: clear value — fixes context budget calibration for V3 pipeline
• Risk: Low — 6 files, +123/-8, well-tested

Posted by github-maintain at $(date -u +%Y-%m-%dT%H:%M:%SZ)