Restore summary fallback for recovered agent threads

[swerner]

Fixes #102.

When two full-fidelity agents share a thread across a command, restarting during the command currently resumes successfully but sends only the second agent's new prompt. Recover the thread's prior use from its recorded events and supply the documented summary:high fallback. Retain the rebuilt conversation so the following agent returns to normal full-fidelity reuse. Previously unused threads still start fresh.

The runtime supplies ResumeInfo as an execution-local capability on both resume paths. This reuses existing recorded events and stage outputs; it adds no persistent transcript store or log format change.

Verification:

• Reproduced on Petri main 56ec3849 and Fabro main ebb1f0bbb (0.387.0-nightly.0), which already includes Preserve workflow context during native skill expansion #118. The uninterrupted control carries QA_FIRST_CONTEXT; the recovered request carries neither that answer nor a summary.
• Rebuilt that same Fabro revision against this patch. Using an isolated local server, scripted Anthropic responses, and a local sandbox, killed only the server with SIGKILL while the intervening command was blocked. Confirmed the command was redispatched, the run succeeded, the second request included QA_FIRST_CONTEXT in its stage summary, and a third agent reused the rebuilt conversation. The uninterrupted control also passed.
• Added binary-level regressions using scripted OpenAI responses for the uninterrupted control, crash during the command, subsequent thread reuse, and a previously unused thread after recovery.
• 19 recovery, compaction, and hooks black-box tests passed; 336 agent, runtime, and driver tests passed. Scoped Clippy with warnings denied, pinned formatting, and whitespace checks passed.

The existing failed-predecessor-after-compaction regression passed separately. Direct-agent resume was not separately retested. The full routine gate, Docker/Daytona tiers, and compatibility corpus were not run. Issue #102 should remain open until this fix lands.