feat: overlay structured training events on metric charts (#271)

[wenzeyu8888-rgb]

Summary

Implements #271: overlay clickable structured training events on TensorBoard metric charts.

Changes

Backend (areno/dashboard/server.py)

• detect_training_events(): read-only derivation of 4 event kinds from existing metrics + logs (no new artifact files, no trainer/metrics changes)
  • non_finite: NaN/Inf in TensorBoard scalars (fixes silent discard at L338)
  • constant_reward: rollout/rewards_std <= 1e-9 with rewards_max == rewards_min double-confirm
  • invalid_batch: rollout/advantages_std <= 1e-6 consecutive streaks (>=3 warn, 1-2 info)
  • oom: OOM pattern in captured subprocess logs (severity=error, no recovery tracking)
• get_job_events(): lightweight read of cached job._events without triggering get_job/_save_state
• GET /api/jobs/<id>/events route inserted before catch-all, with ?types= and ?limit= params
• fetch_events agent tool schema + handler
• Detection injected at end of _load_metric_files to avoid lock contention

Frontend (dashboard/src/)

• eventUtils.js (new): ESM pure functions (step-to-x mapping, event filter, log window, metric context) — testable via node --input-type=module
• EventOverlay: SVG markers (cross/diamond/square/triangle by kind) on chart
• EventFilters: 4 checkboxes independent of metric curve
• EventContextPopover: click-through with log_hint.kind 3-way split — OOM shows +-20 log lines, non-OOM shows metric context (step+-3), none shows metadata only

Tests

• tests/test_dashboard_events_cpu.py: 29 CPU tests covering success/malformed/boundary/overlap/legacy/route-order/injection-position
• Node smoke tests for eventUtils.js

Docs

• docs/reference/dashboard-training-events.rst: API contract, output fields, limitations (500-step truncation, OOM semantics, NaN behaviour change), copyable Python fixture example

Architecture

All 4 event kinds are read-only derived from existing TensorBoard scalars + captured logs at _load_metric_files time. No new artifact files, no new producers, no trainer/metrics changes — strictly per issue constraint: "reuse existing data ... rather than introducing a parallel subsystem".

Test results

• Python: 29/29 PASSED (macOS Python 3.9 + Kaggle Python 3.12)
• Node: 20/20 PASSED
• Frontend build: vite build OK (1831 modules)
• Kaggle GPU: verified end-to-end with fake fixture metrics — events detected, API returns correct results, dashboard UI shows markers + click-through popover

Acceptance criteria

• Test overlapping events, legacy runs without events, click-through to bounded log context, active updates, and no mutation of metric data
• The implementation uses existing AReno contracts and introduces no external database or mandatory sandbox
• Default behavior remains backward compatible
• Focused automated tests cover success, invalid input, and one boundary/failure path
• User documentation includes a minimal runnable example and explains observable output

Closes #271