fix(senpi): harden compaction recovery and Goal migration

[realsigridjin]

Summary

• add map-less GPT-5.6 Sol max reasoning support across capability detection, Shift+Tab cycling, and OpenAI-compatible payload serialization
• recover terminal required-compaction watchdog/truncated-stream failures with a deterministic durable checkpoint while preserving retained-suffix progress and global queued-input chronology
• enforce exact per-turn caps and cooldown admission semantics without weakening hard-limit or generic-failure fail-closed behavior
• migrate legacy budget-limited pi-goal state into Senpi's budget-free builtin Goal store

Verification

• npm run check
• coding-agent affected suite: 13 files / 164 tests passed after rebasing onto current main
• AI reasoning suite: 2 files / 60 tests passed after rebasing onto current main
• final review lanes: code quality, goal verification, hands-on QA, security, context/history, and independent gate all PASS/APPROVE
• installed runtime QA: Senpi 2026.7.30, xhigh/max exposed, queue metadata compatibility preserved

Compatibility

• existing Ctrl+Tab behavior is unchanged; only the supported level sequence expands for eligible map-less Sol models
• manual compaction, generic cancellation, and genuine oversize remain fail-closed
• clearAllQueues() keeps the legacy enumerable { steering, followUp } shape while exposing ordered recovery metadata non-enumerably

Related work

• builds on the terminal queue-restoration foundation merged in fix(interactive): restore terminal compaction input to editor #535
• preserves the cooldown-admission contract discussed in fix(compaction): let admission proceed when the breaker cancelled opportunistic compaction #532

---

Summary by cubic

Hardens compaction recovery with a deterministic checkpoint and preserves queued-input order across UI and native queues; also exposes max for map-less gpt-5.6-sol and migrates legacy budget-limited goals to Senpi’s goal store.

• New Features

  • Expose max for map-less gpt-5.6-sol and send reasoning.effort: "max" via openai-responses, azure-openai-responses, openai-codex-responses, and openai-completions. Shift+Tab now cycles xhigh → max. Explicit maps and null vetoes stay authoritative; non‑Sol variants are unchanged.
  • Auto-migrate legacy pi-goal files into the built-in Goal store; convert budgetLimited to active and drop tokenBudget with no user action.

• Bug Fixes

  • Required compaction recovers from summarization timeouts and upstream_stream_truncated by appending a deterministic checkpoint compaction and rebuilding context under limits; manual and non-target failures stay fail‑closed.
  • Pre-prompt compaction failures now emit willRetry: false so queued input returns to the editor instead of being re-queued against an oversized context.
  • Admission honors the compaction circuit-breaker cooldown: if the breaker cancels compaction, the turn proceeds; other cancellations still fail‑closed.
  • Preserve global enqueue order across steering/follow-up and compaction-owned input; clearQueue() keeps the legacy { steering, followUp } shape and adds a non‑enumerable ordered list for recovery and delivery.
  • Enforce exact per‑turn caps and breaker semantics: ineffective compactions count toward the cap, and blocking summaries aren’t invoked while the breaker is open.

^{Written for commit 503174f. Summary will update on new commits.}