fix: perf batch — model inference bug + 5 perf improvements

[microsasa]

Summary

Fixes a correctness bug in model inference for multi-shutdown sessions, plus five performance improvements identified by the perf-analysis agent.

Changes

Bug fix

• [aw][perf] parser: _first_pass calls _infer_model_from_metrics O(N) times for N shutdown events but only the last result is [Content truncated due to length] #931 — _first_pass inferred session model from the last shutdown's metrics ("last wins"), which gave wrong model for display AND pricing when a session spanned multiple models across resume cycles. Now defers to _build_completed_summary where merged metrics across all cycles are available. Explicit currentModel still takes priority.

Performance

• [aw][perf] render_detail: _render_shutdown_cycles iterates sd.modelMetrics.values() twice per cycle #903 — Merged two separate sum() passes over sd.modelMetrics.values() into a single loop in _render_shutdown_cycles.
• [aw][perf] vscode_parser: _finalize_summary creates redundant dict copies that VSCodeLogSummary.__post_init__ immediately re [Content truncated due to length] #904 — Removed redundant dict() copies in _finalize_summary — was creating dicts from defaultdicts then __post_init__ immediately called dict() again. Now passes defaultdicts directly.
• [aw][perf] vscode_parser.discover_vscode_logs: public API bypasses module-level discovery cache #906 — Made public discover_vscode_logs delegate to _cached_discover_vscode_logs instead of running an uncached glob on every call. Updated all 4 platform tests to mock stat() instead of is_dir().
• [aw][perf] parser: get_all_sessions rebuilds fingerprint and re-sorts on plan.md-only name changes #911 — Added plan-only fast path in get_all_sessions: when only plan.md changed (no events.jsonl modifications), substitutes updated names into existing sorted order instead of rebuilding the O(n) fingerprint and O(n log n) sort.
• [aw][perf] vscode_parser: _update_vscode_summary leaves four accumulator fields as attribute accesses inside the per-request h [Content truncated due to length] #912 — Hoisted total_requests, total_duration_ms, first_timestamp, last_timestamp to locals before the per-request loop in _update_vscode_summary, matching the pattern already used for dict fields.

Won't-fix (closed with explanation)

• [aw][perf] vscode_parser: get_vscode_summary stats every log file on every call before checking the global summary cache #898 — get_vscode_summary stats every log file before checking cache. Benchmarked at 0.1–0.2ms for typical file counts. Skipping stats would risk serving stale data when files are modified — freshness over sub-millisecond savings. Added regression test to guard this.

Testing

• make check passes: lint ✅, types ✅, security ✅, 99% coverage ✅, 86 e2e ✅
• New tests for [aw][perf] parser: _first_pass calls _infer_model_from_metrics O(N) times for N shutdown events but only the last result is [Content truncated due to length] #931: test_dominant_model_wins_over_last_shutdown, test_explicit_current_model_takes_priority
• New test for [aw][perf] vscode_parser: get_vscode_summary stats every log file on every call before checking the global summary cache #898 freshness: test_appended_file_detected_without_discovery_change
• Updated test for [aw][perf] parser: get_all_sessions rebuilds fingerprint and re-sorts on plan.md-only name changes #911: test_sort_runs_after_plan_change (asserts sort is now skipped)
• Updated 4 platform tests for [aw][perf] vscode_parser.discover_vscode_logs: public API bypasses module-level discovery cache #906: mock stat() instead of is_dir()

Review

Reviewed by three adversarial agents (Codex, Opus 4.6, Sonnet 4.6). Sonnet found 3 stale is_dir test mocks — fixed and squashed into #906.

Closes #931
Closes #903
Closes #904
Closes #906
Closes #911
Closes #912

[Copilot]

Pull request overview

Fixes incorrect session model inference for multi-shutdown/resumed sessions and applies several targeted performance optimizations in hot paths (session parsing, VS Code log discovery/summarization, and detail rendering).

Changes:

• Fix model inference for multi-shutdown sessions by deferring inference to merged shutdown metrics in _build_completed_summary (while keeping explicit currentModel highest priority).
• Reduce per-call overhead via multiple micro-optimizations (single-pass shutdown-cycle totals, fewer dict copies when finalizing VS Code summaries, local-variable hoisting in the VS Code aggregation loop).
• Improve caching behavior (public discover_vscode_logs now uses the discovery cache; add a plan.md-only fast path in get_all_sessions to avoid unnecessary sort/fingerprint work).

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
src/copilot_usage/parser.py	Fixes multi-shutdown model inference and adds plan-only cache fast path in get_all_sessions.
src/copilot_usage/vscode_parser.py	Makes public discovery cached; reduces overhead in summary aggregation and finalization.
src/copilot_usage/render_detail.py	Avoids double iteration over per-cycle model metrics when rendering shutdown cycles.
tests/copilot_usage/test_parser.py	Adds/updates tests for multi-shutdown model inference and plan-only sort skipping.
tests/copilot_usage/test_vscode_parser.py	Updates platform discovery mocks to stat() and adds regression test for cache freshness on file append.

Comments suppressed due to low confidence (1)

tests/copilot_usage/test_parser.py:9130

• The docstring for test_sort_runs_after_plan_change still describes a plan.md change forcing a fresh sort via the not deferred_sessions guard, but the updated assertion expects session_sort_key to NOT be called. Please update the docstring and/or rename the test to reflect the new plan-only fast-path behavior.

        sort_key_calls: list[int] = []

        def tracking_key(session: SessionSummary) -> datetime: