perf(coding-agent): cache materialized session views behind mutation counter

[code-yeongyu]

What

Bug A from the render-hotpath plan: the interactive footer re-materialized the entire session 3x per frame at ~60fps (getEntries/getBranch/getSessionName each deep-copying a 10 MB / 724-entry session via ResidentStringStore.materialize). Profiled at 33.4% self-time in transformJsonValue on a 17h session burning 50-86% CPU idle.

Fix

• SessionManager mutation-counter caching: memoized getEntries()/getBranch() (keyed on leafId+mutationCount), O(1) getSessionName()
• Incremental getUsageTotals() maintained on assistant-message append; footer consumes it (legacy loop deleted)
• Robustness: usage-less assistant entries (aborted/error turns in older session files) contribute nothing

Measured evidence (real CLI, 10 MB session, tmux)

• Idle CPU: 50-86% -> 0.0% x5 samples (t3-perf-qa.txt)
• sample(1) top-25 self-time: transformJsonValue ABSENT (was 33.4%)
• footer-width.test.ts byte-identical output guard green
• Cross-fix live QA (F2): first footer render 2.12s, prompt round-trip 0.29s, idle CPU 0.0% x10 over 10 min (f2-live-qa.txt)

Evidence

Plan: .omo/plans/senpi-render-hotpath-and-startup-fixes.md

• t1-vitest.txt / t1-vitest-red.txt (RED->GREEN mutation-counter cache)
• t2-vitest.txt / t2-vitest-red.txt (RED->GREEN usage totals equality)
• t3-perf-qa.txt (manual perf QA, PASS)
• pr1-full-gate.txt (gates + environmental flake characterization: origin/main itself fails 74/284 mcp tests under local parallel load; CI is the authoritative full gate)

Plan: .omo/plans/senpi-render-hotpath-and-startup-fixes.md

[code-yeongyu]

Todo 2 (incremental usage totals) landed in 40d98ec: SessionManager.getUsageTotals() maintained incrementally on assistant-message append, rebuilt in _buildIndex()/newSession; footer hot path no longer scans getEntries() per frame. TDD: totals-equality test (legacy loop vs getUsageTotals(), deep equality incl. latestCacheHitRate + all-entries semantics) captured RED then GREEN; footer-width.test.ts green with zero output diffs (125/125 tests passing across footer-width + session-manager suites). Evidence: .omo/evidence/senpi-render-hotpath-and-startup-fixes/t2-vitest-red.txt and t2-vitest.txt.

[code-yeongyu]

CI failure root cause + fix (commit 9bc68d9)

The 3 CI failures (footer-token-format, startup-session-name x2) traced to two gaps from the incremental-usage-totals change:

1. session-manager.ts _accumulateUsage dereferenced entry.message.usage unconditionally. Older session files contain assistant entries with no usage (aborted/error turns), which crashed on load and broke session-start paths (startup-session-name). Fixed: early-return when usage is undefined — those entries contribute nothing to the running totals.
2. footer-token-format.test.ts mocked sessionManager without the new getUsageTotals method the footer hot path now calls. Fixed: mock returns totals matching the fixture transcript.

Verification

• Targeted vitest (green, 16 files / 133 tests): test/footer-token-format.test.ts test/startup-session-name.test.ts test/session-manager test/footer-width.test.ts test/suite/app-server-daemon.test.ts
• Pre-commit gate (npm run check: biome, tsgo, shrinkwrap/install-lock, neo build+vet+test) green on the commit hook.

Environmental note (local full-suite flakiness, pre-existing)

Local full-suite runs on a loaded parallel machine flake in spawn-heavy suites unrelated to this PR (app-server daemon/stdio CLI spawns, resource-loader, mcp). Confirmed pre-existing: plain origin/main itself fails 74/284 test/mcp tests under the same local parallel load. One earlier local failure was also caused by a leaked daemon from a killed run holding port 18999. Per vitest.config.ts, the serialized CI pool is the deterministic configuration, so GitHub CI is the authoritative full gate for this PR.

Evidence: /Users/yeongyu/local-workspaces/.omo/evidence/senpi-render-hotpath-and-startup-fixes/pr1-full-gate.txt

[code-yeongyu]

CI failures fixed (commit 9bc68d9)

Root causes — two distinct issues behind the 3 failing tests:

1. test/footer-token-format.test.ts — its sessionManager mock was still built around getEntries() and lacked getUsageTotals() after Todo 2 moved footer.ts to sessionManager.getUsageTotals(). Fixed by adding a getUsageTotals mock (same legacy-formula pattern as test/footer-width.test.ts).

2. test/startup-session-name.test.ts (both tests) — not a mock problem; a real production regression. SessionManager._accumulateUsage() assumed every persisted assistant message has usage, but session files can contain usage-less assistant entries (e.g. aborted/error turns — the test fixture writes one). The legacy footer loop only ran in interactive mode, so these files loaded fine before; the new load-path accumulation crashed with "Cannot read properties of undefined (reading 'input')" during print-mode startup, so --name never got appended. Fixed with a minimal guard in _accumulateUsage (skip usage-less entries), matching legacy semantics.

Gate result: full gate green from worktree root — npm run check && CI=true npm run test exits 0. coding-agent: 462 passed / 4 skipped (466 files), 4006 passed / 31 skipped (4037 tests); all other workspace packages pass. (CI=true selects the repo's serialized forks pool per vitest.config.ts guidance for subprocess-heavy suites.)

Note: local reruns were hampered by extreme machine load (loadavg ~180) which flaked pre-existing subprocess tests (app-server-daemon, app-server-stdio, skills-sidecar) — those pass standalone and are unrelated to this change.