feat(storage): global trash cleanup, reference lists and browser profile status (C22–C24)

[yxflc11]

What changed

Implement the C22–C24 backend contracts from docs/research/storage-cleanup-follow-ups.md:

• C22: one Owner-only global trash cleanup with a receipt keyed only by requestKey, reusing C21 deletion/recovery/reference guards and audit. Storage totals include measured referenced trash bytes. Migration 0051 adds global receipts (52 migrations total).
• C24: bounded Owner-only attachment reference lists using C19 matching, with short previews, exact counts and owner/bot/system authors. The Owner explicitly approved system authors.
• C23: carry safe-integer/null profileBytes through browser maintenance and the Docker Provider. The previously requested upstream contribution is tracked in feat(agent-computer): report bounded per-Bot profile bytes CopilotKit/OpenBot#730. The deployed dependency pin stays unchanged, so existing images return null until the reviewed route is available.

UI wiring is a separate consumer change. This PR targets Peerframe/openbot only and does not enable automatic merge.

Acceptance journey

1. An authenticated Owner clears trash across active channels; referenced files remain, each deletion is audited, and retrying the same requestKey touches no later files.
2. The Owner requests attachment references, including recycled files, and receives bounded same-channel message/task lists. Missing authority and purged files withhold content.
3. Browser status carries measured profileBytes when available and null for unsupported/refused measurements.

Open-source research

• Research artifact: docs/research/storage-cleanup-follow-ups.md; existing channel-storage-purge.md, channel-attachment-reference-counts.md and desktop-browser-management.md decisions.
• Selected upstream/standard: existing PostgreSQL and C21/C19 contracts; agent-computer authenticated HTTP surface; Node filesystem APIs and Linux open(2).
• Version or commit: deployed agent-computer 257c1280d684089be9adb0b35cce262efc7064bf; reviewed contribution base cb5dc32a44517622c6db4e527e61d3abb389b43c, contribution head 46eb7af817027c5de4202846c73c43bbb2fa67b7; PostgreSQL 17.11 qualification.
• License: MIT for agent-computer and local implementation; existing PostgreSQL license.
• Decision: extend existing control authority and adapters; retain the deployed dependency pin while the prior upstream contribution is pending. No runtime fork adoption.
• OpenBot-specific gap: globally keyed cleanup receipts, per-attachment reference previews and nullable per-Bot profile measurements.
• Source copied or substantially adapted: no.

Verification

• Focused real PostgreSQL/HTTP tests: 52 passed; protocol: 3 passed; Docker browser/maintenance: 35 passed.
• Migration qualification: 40 retained upgrade/restore cases and 8 cleanup cases passed; exact committed-source evidence in experiments/s7-migration/evidence/global-trash-result.json.
• Actual Linux arm64 product image build/smoke passed with all 52 migrations, Owner HTTP, built Web, parsers and restart persistence.
• Broader completed checks include Web 644 tests, Docker Provider 67 tests, Node 129 Vitest tests (3 skipped) plus 54 transport tests, and the final build (18 tasks, 13 cached).
• npm run check did not finish green: an unchanged publisher CLI timed out, then a serial test retry hit four Desktop timeouts. Publisher tests passed alone; all 15 tests in the three Desktop files passed with one worker. Remaining Node tests and build were completed separately.
• npm run test:control:python exceeded the existing 300-second suite deadline twice. The second completed 789 tests without failures before timeout. Remaining modules were supplemented with disposable SQL tests (301 completed; 13 initially failed due to omitted diagnostic fixture identities, then all 13 passed after fixture correction). This does not qualify the default aggregate command or its post-pytest TS read-back assertions.
• Documentation checks passed; research unit tests passed. Hosted CI and dedicated Worker/Temporal qualification remain pending/unrun. Full details and prior failures are retained in the bilingual research record.
• Manual path: no production mutation, paid model call or UI acceptance claimed.
• Platforms actually tested: macOS arm64 and local Docker Linux arm64.
• Support level claimed: backend contracts implemented; C23 numeric production measurements and UI integration remain dependent work.

Security impact

• New cleanup surface documented with fail-closed tests; authority remains with the Owner/control plane.
• No credentials or private user data included.

Documentation and compatibility

• Implemented, pending and planned behavior are distinguished.
• English API/research/reuse documents and Chinese translations updated.
• Platform boundaries and upstream measurement limitations documented.

AI or automation assistance

• Tools/models used: Codex, existing repository tests, disposable Docker fixtures.
• Human-reviewed areas and evidence: Owner-approved contracts and system author addition; implementation awaits PR review.