Preserve entry timestamps on compaction and branch-summary replay messages

[baixinghao]

Summary

Fixes #747.

SessionState.from_entries builds a synthetic UserMessage for compaction and branch_summary entries without passing timestamp=, so UserMessage's default_factory=current_timestamp_ms fills in wall-clock "now" at replay time instead of the entry's historical entry.timestamp.

This breaks _last_applicable_provider_usage (src/tau_coding/context_window.py), which anchors token accounting on the latest assistant response's provider-reported usage but rejects any usage older than a later message. Because the replayed summary message's timestamp is always "now", every assistant usage after a compaction fails the ordering check and context accounting falls back to the crude 4-chars-per-token estimate for the rest of a compacted-and-reloaded session.

Fix

Pass timestamp=round(entry.timestamp * 1000) at both sites, mirroring the custom_message fix from #710:

• _apply_compaction's summary_row
• the branch_summary case in SessionState.from_entries

Verification

Reproduced the issue's script on this branch:

• Before: summary UserMessage timestamp= 1791756578906 (wall clock), estimate_context_usage returned uses_provider_usage=False, provider_tokens=0.
• After: UserMessage timestamp= 1700000010000 (entry time), uses_provider_usage=True, provider_tokens=999.

Added test_compaction_and_branch_summary_messages_keep_entry_timestamps covering both the compaction summary row and the branch-summary row.

uv run pytest tests/test_session.py — 42 passed. uv run ruff check / ruff format --check / mypy on the touched files — clean.

Note: this Windows environment has ~40 pre-existing test failures on a clean checkout (TUI layout, updater, file-drop path handling, etc.). I diffed the failure list before/after this change: identical, so the change introduces no regressions.