feat(ego-browser): reclaim idle agent task spaces to release renderers

[Vitus213]

What

Add lazy idle task-space reclamation so agent task spaces left by earlier sessions — and their renderer processes — get released automatically instead of piling up.

Closes #270.

Why

Today completeTaskSpace is the only release path and it is 100% agent-driven; abandoned sessions leak renderer processes and swap climbs monotonically (see #270 for 26h measurements: 20→58 helper processes, 10→35 GB swap). The harness kept no cross-process activity state and listTaskSpaces carries no timestamp, so there was nothing to time out against.

How to verify

cd package/ego-browser
npm ci
npm test    # 312 pass / 0 fail

Behavior covered by the new taskspace-reclaim.test.mjs:

• a space idle past EGO_RECLAIM_IDLE_S (default 7200s) is closed; a fresh one is kept
• ownership === "user" is never touched
• hard cap (EGO_RECLAIM_MAX_SPACES) reclaims oldest-first when over the cap
• EGO_RECLAIM_DISABLE=1 is a no-op
• first-sight seeding: spaces with no activity record are seeded with now and skipped (deploy safety — never mass-closes pre-existing / in-use spaces); they only become reclaimable after the idle window
• a seeded space is reclaimed on a later run once idle
• no ego runtime → empty result, no throw

Manual (against a live browser):

ego-browser nodejs <<'EOF'
const t = await taskSpaces.useOrCreate('reclaim-verify');   // triggers the once-per-process reap
console.log(JSON.stringify({ id: t.id, spaces: (await listTaskSpaces()).length }));
await taskSpaces.complete('reclaim-verify', { keep: false });
EOF

Impact

• Helper surface: no new public helper. Reclamation is an internal mechanism triggered from run.ts execute() (CLI mode) and helpers.useOrCreateTaskSpace() (embedded app-SDK mode), collapsed to once-per-process via state.reclaimDone.
• Agent side: no change required. Agents that already call completeTaskSpace behave identically; agents that forget it now have a safety net.
• Existing tests: taskspace-e2e.test.mjs and helpers.test.mjs set EGO_RECLAIM_DISABLE=1 in their harness so call-sequence assertions stay deterministic. New suites: taskspace-activity.test.mjs, taskspace-reclaim.test.mjs.
• Config: env-gated, default on (idle 2h, cap 8). EGO_RECLAIM_DISABLE=1 opts out.

Files

• src/taskspace-activity.ts (new) — activity map + atomic persistence + config
• src/taskspace-reclaim.ts (new) — reclaim logic; talks to globalThis.ego directly (no helpers import → no cycle); once-per-process guard
• src/helpers.ts — selectTaskSpace touches activity (gated by disabled); useOrCreateTaskSpace triggers reclaim (CLI+app-SDK coverage)
• src/run.ts — execute() triggers reclaim (CLI coverage)
• src/state.ts — reclaimDone once-guard
• AGENTS.md — Task Spaces section documents the mechanism + env

Open questions for reviewers

1. Are the default thresholds (2h idle / 8 cap) right for the project, or should reclamation be opt-in by default?
2. useOrCreateTaskSpace as the app-SDK trigger point — acceptable, or prefer a different hook?

[Vitus213]

@section9-lab @WUXM5 Hi — this PR has been open for a few days with no review. It reclaims idle agent task spaces so their renderer processes are released, cutting memory footprint when agents park between rounds. Happy to adjust scope or add tests if needed — just let me know what would help move this forward. Thanks!