feat(dashboard): show delayed node eligibility time

[breken-ai]

Closes #611

What goes wrong

When a node is delayed (ReQueueAfterSignal / RequeueAtSignal, a retry-policy backoff, or a graph start_delay) the state manager stores the moment it becomes eligible in State.enqueue_after (epoch milliseconds; enqueue_states picks states with enqueue_after <= now). That value never reaches the dashboard: NodeRunDetailsResponse does not carry it, so the Node Details modal shows a CREATED node with no hint of why it is not running or when it may.

What changes

state-manager

• NodeRunDetailsResponse.enqueue_after: Optional[int] = None (epoch ms), described as the stored eligibility time and explicitly not a guaranteed execution time.
• get_node_run_details passes state.enqueue_after through verbatim (stored value, not recomputed).

dashboard

• NodeRunDetailsResponse.enqueue_after?: number | null in the client types.
• New pure helper src/lib/nodeEligibility.ts: getNodeEligibility(details, nowMs) returns scheduled (future) or waiting (past-due, still not picked up) only for CREATED states with a usable timestamp; null otherwise. formatEligibilityTime renders the instant in the viewer's locale/timezone with a zone label. Wire representation stays epoch ms; nothing localized is ever sent back.
• NodeDetailsModal shows, under Status & Timestamps:
  • Eligible after: <local time TZ> — "Delayed by a start delay, retry policy or requeue signal. A worker can pick this node up any time after this moment; it is not a guaranteed start time."
  • Eligible since: <local time TZ> — "Eligible to run and waiting for a worker to pick it up."
    The clock reading is paired with the details object it was taken for, so a node opened long after page load is never classified against a stale reading. A 5 s interval flips scheduled → waiting on time; the timeout and interval are cleared whenever details change or the modal unmounts.
• npm test script (node --test, zero new dependencies) for the helper; allowImportingTsExtensions in tsconfig.json (valid because the project is noEmit) so the test can import ./nodeEligibility.ts the way Node's type-stripping loader requires.
• Docs: one paragraph in docs/docs/exosphere/dashboard.md and a feature bullet in dashboard/README.md.

What deliberately does not change

• No scheduler/dispatch logic: enqueue_states, the requeue signals and the retry path are untouched. This is observability only.
• No new persisted fields; enqueue_after already exists on every State document.
• Nothing is shown for QUEUED/EXECUTED/SUCCESS/ERRORED/… states, so a node that was already picked up never carries a stale "waiting" message.
• Records or servers without the field (undefined/null/0) show nothing — no Invalid Date, no 1970 epoch.
• Existing tests were not modified; the existing MagicMock fixtures still pass because the field is optional.

Tests

• state-manager: tests/unit/controller/test_get_node_run_details.py +2 (exposes_stored_enqueue_after, response_enqueue_after_is_optional). Full suite 639 → 641 passed.
• dashboard: src/lib/nodeEligibility.test.ts 0 → 7 passed (fake clock via injected nowMs, no real timers): future delayed node, past-due node, boundary at == now (matches the server's <=), missing/null/0/NaN field, every non-CREATED status, null/undefined details, and the same instant rendered in UTC / Asia/Tokyo / America/Los_Angeles across a day boundary.

Both new tests fail on main (AttributeError: 'NodeRunDetailsResponse' object has no attribute 'enqueue_after'; ERR_MODULE_NOT_FOUND for the helper).

Evidence record

Repository / issue: FailproofAI/runtime — #611 "feat: Show next run time of state"
Upstream SHA / patch SHA: 2db4b154c8bb020051f6ba349ed922f8479eef13 / fc5e6c9ef845f6deea883756a5cdceffcfcaca32
Issue ownership and competing PR result (2026-09-12T21:49:28Z UTC):
  state=open, assignees=[], updated_at=2026-01-03T14:58:35Z; only comment is the CodeRabbit plan bot;
  cross-references: #616 (state querying, unrelated feature issue);
  gh search prs "611" -> none; gh search prs "enqueue_after next run time" -> none;
  open PRs (37) -> none touch node run details / enqueue_after.

Reproducer command and red result:
  state-manager$ uv run pytest tests/unit/controller/test_get_node_run_details.py -q
    -> 2 failed, 6 passed  (AttributeError: 'NodeRunDetailsResponse' object has no attribute 'enqueue_after')
  dashboard$ node --test src/lib/nodeEligibility.test.ts
    -> fail 1  (ERR_MODULE_NOT_FOUND ./nodeEligibility)

Patched command and green result:
  state-manager$ uv run pytest tests/ -q          -> 641 passed (baseline on main: 639 passed)
  dashboard$ npm test                             -> tests 7, pass 7, fail 0

Full checks executed (all local, macOS, Python 3.12 via uv, Node 25.6.1):
  state-manager$ uv sync --group dev
  state-manager$ MONGO_URI=mongodb://admin:password@localhost:27017 MONGO_DATABASE_NAME=test_exosphere_state_manager \
      STATE_MANAGER_SECRET=test-secret-key SECRETS_ENCRYPTION_KEY=<CI value> uv run pytest tests/ -q
      (mongo:7 in Docker with the CI credentials)          -> 641 passed
  state-manager$ uv run ruff check <3 changed .py files>   -> All checks passed!
  dashboard$ npm ci
  dashboard$ npx tsc --noEmit -p tsconfig.json             -> exit 0
  dashboard$ npm test                                      -> 7/7
  dashboard$ npx next build                                -> ✓ Compiled successfully, exit 0
  dashboard$ eslint on the 4 touched files with typescript-eslint recommended + react-hooks recommended
      (ad-hoc flat config, not committed)                  -> 0 findings on changed lines; the 2 reported
      errors (react-hooks/set-state-in-effect) are on pre-existing lines of NodeDetailsModal.tsx.

Checks not executed and why:
  dashboard$ npm run lint  -> cannot run on upstream main: `next lint` was removed in Next 16
      ("Invalid project directory provided ... /dashboard/lint"), and `eslint src` with the repo's
      eslint.config.mjs fails at config load (circular structure in FlatCompat + eslint-config-next 16;
      eslint-config-next's bundled eslint-plugin-react also calls context.getFilename, removed in ESLint 10).
      Both are pre-existing and unrelated to this change; not fixed here to keep the PR in scope.
  No browser-driven test of the modal: the dashboard has no component test runner; the display rule is
  isolated in the pure helper and covered there. Verified visually by reading the JSX only.

https://claude.ai/code/session_01DkE5qM85ht9Uoq3aAqcKf7