feat: show structured cycle progress and confirm stop cleanup

[MaxMiksa]

The dashboard previously depended on free-form final reports, so active cycle progress, actual execution evidence and model configuration were often unavailable. This change adds validated cycle work reports and deterministic Codex event collection while keeping runtime outcomes, model statements and registered checks separate.

• Show concise per-cycle titles, progress, blockers and recorded next actions, with explicit missing/invalid/interrupted fallbacks. Reuse /api/journal; no additional model calls for extraction.
• Capture bounded command/file events and exact-session model settings. Provide opt-in check, document and static-preview registration with freshness and ownership checks.
• Confirm Windows/WSL stop cleanup, preserve interrupted-cycle evidence, and expose failed cleanup for retry.
• Include the previously prepared bilingual usage guidance, current dashboard screenshot and side-by-side product examples.

Validation:

• Six actual model calls: three continuous cycles each with Luna/high and Terra/high. Exact session configuration confirmed for every cycle; 12 distinct work reports and 264 live HTTP samples. Prior reports, per-cycle evidence, selection, product language and Human Overrides remained intact.
• Local Python: 274 executed and passed, 8 live-platform cases skipped for the separate hosted service suites.
• Frontend semantics: 8 passed. Existing Chromium suite: 19 passed. Two additional browser checks used the retained real three-cycle histories.
• Final candidate CI 35417724406 passed all 9 selected jobs, including the required gate and real isolated systemd/macOS launchd suites. The 2 unchanged example jobs were skipped by path routing. The first run exposed a missing module in the browser fixture; its source copy list is now fixed, with no change to production behavior or weakened assertions.

Scope: Codex runtime observations were verified with CLI 0.155.0. Work reports remain model-authored descriptions; recorded next actions do not assert execution, and passing checks do not imply complete business acceptance. Other engines retain their existing behavior and report unavailable observations explicitly.