spb/forge-studio Public

- ChartXState shared between the main chart and all secondary charts:
  zooming/panning up top moves every chart, and the hover crosshair is
  synchronized - secondary headers show the value AT the cursor
- Checkpoint annotations on the loss chart (green dashed rules with a
  drive glyph); clicking one selects that checkpoint in the panel below
- New Run form now covers EVERY Forge config field: attention/positions
  (rope theta/global/scaling, NoPE, head_dim, bias, softcap), norm
  placement, QAT, the full MoE section (V3 sigmoid/noaux/scaling/shared/
  first-k-dense), advanced optimizer (Muon, WSD, betas, forge_save/dtype)
- Export menu: chart as 2x PNG (ImageRenderer), metrics as CSV
- Run queue: starting a run while one is live persists it as "queued";
  the supervisor launches the oldest queued run when the active one
  exits - trainings never overlap, nothing is lost

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- CompareView: 2-8 run overlay (val + optional train-EMA), X axis in
  steps / TOKENS (honest cross-batch-size axis) / wall-clock from
  elapsed_s, auto hyperparameter-diff legend (only differing fields),
  sortable summary table (params/best val/final ppl/mean tok-s/duration),
  CSV export
- DatasetsView: token counts (bin headers), vocab, on-disk size; New
  Dataset sheet drives tools/prepare_data.py (TinyStories) or
  tools/prepare_hf_data.py (HF sources + research presets) with a live
  streamed console
- Sidebar restructured: Studio (Comparer, Datasets) + Runs
- Local notifications on run finish/fail (final loss / reason in body)
- Generated app icon (forge-gradient + flame, full iconset -> icns),
  wired into the bundle
- Stress test: 200k-line CSV ingest ~1.1s, LTTB+EMA snapshot <250ms —
  documented in RESEARCH.md §9

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

- RESEARCH.md: ground-truth Forge contract (config schema incl. waves 1-3,
  CLI, log.csv grammar with elapsed_s, checkpoint/resume, signals)
- ForgeConfig: byte-compatible Codable mirror of every model/train field,
  param-count formula cross-checked against forge info, LR-schedule math
- Services: ProcessRunner (actor, line streaming), LogParser (header-driven
  CSV + stdout events), MetricsStore (actor, incremental CSV tail, LTTB
  snapshots), RunStore (atomic registry), RunSupervisor (state machine,
  launch/stop, 1 Hz ingest), ForgeBinaryLocator, SystemInfo
- Charts: LTTB downsampler, bias-corrected EMA
- UI: NavigationSplitView shell, run rows/badges, RunDetail dashboard
  (loss raw+EMA+val, log-Y, best-val marker, LR/tok-s/grad-norm secondary
  charts, console), New Run sheet (presets, derived panel, LR preview,
  inline validation), Settings with forge-info validation
- scripts: package-app.sh (SPM -> .app, ad-hoc) and notarize.sh with the
  real zyquo-term identity (Team 3YM54G49SN, profile MacLustr-Notarize)
- 8 tests green: config round-trip vs real configs, param count, LTTB,
  EMA, CSV/stdout parsers, state machine, LR preview

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>