Emit and render threshold and manual compaction events

[baixinghao]

Summary

Fixes #506.

Threshold-based automatic compaction ran silently: when the context estimate crossed the auto-compact threshold, CodingSession issued the summarization request at a turn boundary with no events at all, so the TUI showed a finished answer followed by an unexplained spinner. Only the overflow path emitted CompactionStartEvent / CompactionEndEvent, even though CompactionReason already defines "manual" | "threshold" | "overflow".

Changes

src/tau_coding/session.py

• Split _maybe_auto_compact into _auto_compact_plan() (threshold + plan checks) and _auto_compact_events() (execution). The latter is an async generator that yields CompactionStartEvent(reason="threshold") and CompactionEndEvent(reason="threshold", ...) around a real compaction run, and also forwards them to the extension runtime — mirroring how the overflow path both yields and emits. No plan means no events, so routine threshold checks stay silent.
• A swallowed summarization exception now also yields CompactionEndEvent(reason="threshold", aborted=True, error_message=...), so failures are observable instead of only appearing in the diagnostic log.
• All three auto-compact call sites (auto_compact_before_prompt, auto_compact_after_prompt, auto_compact_after_continue) forward the events into the prompt/continue streams. Compaction stays inside the enclosing turn: no extra AgentSettledEvent, so no extra turn-finished notification.
• compact_detailed() (manual compaction) now emits CompactionStartEvent(reason="manual") / CompactionEndEvent(reason="manual", ...) through the extension runtime, including an aborted end event on failure, so extensions and non-TUI frontends get one consistent signal for every compaction kind. The TUI's own /compact flow is unchanged.

TUI rendering (tui/adapter.py, tui/app.py)

• A … Compacting context transcript status item explains the post-answer wait while threshold compaction runs.
• A failed threshold compaction shows … Auto-compaction failed: <error> instead of passing silently.
• Overflow rendering is untouched.

Docs (website/content/guides/context.md): document the new status items.

Tests

• test_threshold_auto_compaction_emits_exactly_one_start_end_pair — one adjacent start/end pair with reason="threshold" across the prompts that trigger compaction, and exactly one AgentSettledEvent per prompt.
• test_auto_compact_check_without_work_emits_no_compaction_events — threshold checks that find no work emit nothing.
• test_threshold_auto_compaction_failure_emits_aborted_end_event — a failing summary provider yields an aborted end event with the error message, the turn still completes, and nothing is persisted.
• test_manual_compaction_emits_events_for_extensions — manual compaction emits one start/end pair with reason="manual" and the result attached.
• test_tui_adapter_renders_threshold_compaction_status — status item on start, silence on success, failure status on aborted end.
• test_tui_prompt_worker_explains_threshold_compaction_pause — end-to-end prompt-worker flow: answer renders, compaction status explains the pause, running state and notification bookkeeping stay owned by the turn.

uv run pytest tests/test_coding_session.py tests/test_tui_adapter.py tests/test_tui_app.py passes; ruff check / ruff format --check / mypy are clean on the touched files.

Note: this Windows checkout has ~40 pre-existing test failures on a clean main (TUI layout, updater, file-drop path handling, etc.). The failure list before/after this change is identical — no regressions.