feat(compaction): stream the summary into the chat as it generates

[hakula139]

Summary

/compact and auto-compaction used to show only a spinner while the summary was being generated, which could take tens of seconds in a long session. Now the summary streams into the chat as it is generated, using the layout of the compacted block it will become. Every compaction outcome also removes or replaces it cleanly.

• Protocol: Two new events, Event::SummaryToken(String) and Event::SummaryDiscarded, carry summary updates. A sequence of SummaryToken chunks ends with SessionCompacted on success or SummaryDiscarded otherwise. A discard precedes the TurnFailed, Cancelled, or Error event that reports the outcome. If an overflow triggers a retry, the discard comes before streaming resumes, so the frontend never shows two attempts at once. Both events are included in the every_event round-trip table.
• Engine: compact_session takes the event sink and reads the reply through a SummaryStream sink, which forwards text as SummaryToken, passes retry notices through, and drops thinking. It emits SummaryDiscarded only if text has streamed. Its Drop implementation handles cancellation, since await_unless_aborted drops the summarization future before the caller reports Cancelled. replace_session_with_summary discards the summary before reporting a failed boundary write.
• TUI: ChatView keeps the in-flight summary in its own buffer and displays it a blank row below the chat using CompactedBlock::render_streaming, under a Summarizing the conversation... header. SessionCompacted replaces it with the final block. SummaryDiscarded removes it and subtracts it from the paused new-content count.
• Stdio: The REPL and headless frontends ignore both events and keep their one-line boundary notice. Headless runs never auto-compact because the engine starts without usage to trigger it.
• Docs: The /compact design, including its component notes and design decision 12, the auto-compaction steps, and the welcome screen's emptiness rule now describe the stream.

Closes #186

Test plan

• cargo fmt --all --check
• cargo build
• cargo clippy --all-targets -- -D warnings
• cargo test: 2659 unit and 4 CLI tests pass.
  • Compaction tests assert the exact event sequence for the initial text, thinking, overflow retries, prompt-too-long, the newest-round overflow, empty and unfinished summaries, retry notices, and a future dropped mid-summary.
  • Auto-compaction tests assert the full sequence for an overflow retry and an unsaved boundary. Engine tests cover manual success, failure, a dead session, and cancellation after the first token, where the discard arrives before Cancelled.
  • TUI snapshots cover streaming, completed, and cancelled states, and a text test covers a failure.
• cargo llvm-cov: Every changed line is covered except the closing line of the test server task in serve_stalled_summary, which keeps its connection open by never finishing.
• pnpm format, pnpm lint, pnpm spellcheck

[codecov]

Codecov Report

❌ Patch coverage is 99.77629% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
crates/oxide-code/src/engine.rs	98.63%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!