perf(agent-org): reuse bounded time formatters

[ShiboSheng]

Problem

Fixes #765

Repeatedly switching between the Agent Org Group projection and individual Member pages rebuilt browser-local Intl.DateTimeFormat / ICU formatters during rendering. The Group activity rows, Member turn-page range, Member timing labels, and discussion events bypassed the existing bounded formatter cache, creating avoidable WebKit allocation churn across remounts.

The remaining ps RSS increase is not retained dirty memory: macOS footprint classifies more than the RSS delta as reclaimable JavaScriptCore pages while physical footprint and dirty WebKit malloc both fall below baseline. This PR fixes the real formatter churn and records both metrics rather than claiming that raw RSS immediately returns pages to the OS.

Solution

Add short local-time helpers backed by the existing 64-entry Intl.DateTimeFormat LRU and route every affected Group/Member hot path through them. The helpers preserve the prior locale inputs and formatting options, including the Member page's 24-hour output.

The resulting invariant is: Agent Org render paths must reuse the shared bounded date formatter instead of constructing locale formatters per render. Regression tests cover the Group projection owner and prove that repeated Member page/timing formatting does not call Date#toLocaleTimeString or construct additional Intl.DateTimeFormat instances.

Potential risks

• Timestamp text is intended to remain unchanged; output-parity tests cover the existing locale/options combinations, but other operating-system locales were not manually enumerated.
• A formatter resolved from the operating system's default locale remains cached until eviction or process restart. This matches the repository's existing bounded formatter behavior but means an in-process OS locale change is not reflected immediately.
• Raw WebContent RSS still rises because JavaScriptCore keeps reclaimable heap pages. The physical footprint, dirty WebKit malloc, and hidden lifecycle gates pass, but this change does not promise immediate OS reclamation of RSS pages.
• Verification used the macOS BuildFast package. Windows/Linux runtime behavior and release-only signing, notarization, updater, and installer paths were not exercised because this is a frontend-only allocation fix.
• There are no persistence, schema, IPC, wire-format, dependency, or data changes. Rollback is a normal revert of this commit; no data recovery is required.

Verification

• pnpm exec vitest run --config config/vitest.config.ts src/util/data/formatters/dateLocalDisplay.test.ts src/engines/ChatPanel/ChatHistory/utils/__tests__/turnPageFormatting.test.ts src/engines/ChatPanel/ChatHistory/utils/__tests__/turnTimingFormatting.test.ts src/engines/ChatPanel/ChatHistory/utils/__tests__/turnFormattingIntlCache.test.ts src/engines/ChatPanel/ChatHistory/GroupChatView/__tests__/AgentOrgGroupProjectionView.test.ts — 5 files, 29 tests passed.
• pnpm test -- --reporter=dot — 1,336 files and 10,516 tests passed.
• pnpm run typecheck — passed.
• pnpm run lint — oxlint and ESLint passed with zero errors/warnings.
• pnpm run check:circular — no circular dependencies across 6,426 modules.
• pnpm run check:test-placement — passed across 445 directories.
• ORGII_AGENT_ORG_REDESIGN=1 pnpm run tauri:build:fast — passed; webpack runtime 49121; packaged src-tauri/target/dev-build/bundle/macos/ORG2.app in 494.9 seconds.
• Packaged binary SHA-256: 862e7dc7656bb15081e1328b01620662717c581f7c0570da0a75d614d6bbe2d6.
• Computer Use operated the packaged app against the real Flappy Bird Agent Org run: 2 preheat cycles, 20 Group → Planner → Group cycles, 5 minutes foreground idle, and 5 minutes hidden. The final UI remained on Group view with all 18 activity timestamps visible.
• Visual screenshots are not useful for this change because the timestamp output is intentionally unchanged; the packaged-app lifecycle and memory measurements are the acceptance evidence.

Performance evidence

Stage	WebContent RSS	Physical footprint	Dirty WebKit malloc	Reclaimable	Dirty graphics
After 20 cycles vs. post-preheat baseline	+161.1 MiB	-80.8 MiB	-86.8 MiB	+244.7 MiB	0 MiB
After 5 minutes foreground idle vs. baseline	+196.1 MiB	-84.7 MiB	-86.9 MiB	+347.8 MiB	-3.7 MiB
Hidden minute 5 vs. hidden minute 1	+1.7 MiB	-0.9 MiB	-1.6 MiB	+6.5 MiB	+0.8 MiB

No polling, timer, subscription, worker, persistence, or cache-lifecycle mechanism was added. The hidden process showed no staircase growth after its transition.