0.3.3: stable release with the integrated documentation set - #58

- chronos-ui / chronos_app.py start the rewritten Web UI (FastAPI + single-page
  app in chronos/webui, logic in chronos/workbench); --legacy keeps Gradio.
- zh-Hans / zh-Hant / en / ja, browser-language detection, hover help after 2 s;
  buttons, chips, options and picker places are translated.
- Path fields: "Browse..." lists the server's folders; "Upload..." opens the
  native dialog on the server's own machine (path used in place) or uploads to
  <workdir>/uploads/ from another machine. Several datasets merge into one JSONL.
- --lan (off by default): key from CHRONOS_UI_KEY or generated; the server's own
  machine needs no sign-in. --max-upload-mb, --no-native-dialog.
- chronos lazy-probe and per-validation lazy-expert probe in training runs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

On Linux CI MLX is absent, so resolve_backend("mlx") falls back to cpu and the
MLX variants dedupe into the CPU set. The test now resolves backends as
requested, and checks the fallback collapse separately.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The chronos kernel now trains general models with no translation-specific
code. Everything that builds the en<->zh benchmark model lives in
demos/translation (package chronos_translate, CLI chronos-translate):
corpus pipeline, BLEU/chrF runs, doc-mix runs, FLORES+ probe, model
bundles, the 50M recipe and the doc-mix plan. It imports the kernel, is
not part of the Project_Chronos wheel and is not published on PyPI.

Kernel changes:
- LongRun task hooks (task_eval, eval_samples_for, generate,
  write_samples, pinned_modules); task-eval errors are logged as
  task_eval_error and never stop a run.
- load_spec(spec_cls, extra_path_keys) resolves paths against the recipe
  folder and rejects unknown keys with a pointer to the task runner.
- Tokenizer defaults carry no language tags; special_tokens_with() adds
  task tokens. Shard readers move to chronos.data.shards.
- chronos corpus, chronos bundle and the translate extra are removed.

Monitoring of the lazy-expert probe:
- The monitor gets a Lazy experts tab (four tables, history charts, probe
  log) and GET /api/lazy_probes.
- The Web UI probe card shows every measured field in all four languages.
- The run snapshot carries all probe fields plus the full-memory
  reference; a stray "null" on the Web UI monitor page is fixed.

CI also runs on pull requests into dev.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

… limit

- Router: padding mask, telemetry over all micro-batches, lazy decode
  uses the grouped kernel (fidelity 1.0 at every budget).
- Cluster locality: cluster_locality_mode="segment" learns domain
  clusters per document from content alone (default stays "window").
- MLX: cache limit min(RAM/8, 16 GiB) applied in the kernel; recipe/env
  can only lower it. Multimodal fusion port with Torch parity and
  checkpoint round-trip; pretrain/SFT accept multimodal batches.
- Tools: chat template renders <tools>/<tool_call>; chat_records writes
  per-token .lmask supervision read by SFTRecords; serve passes tools.
- Monitor/WebUI: progress-based ETA per phase; stopped runs show their
  state instead of a clock ETA.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

np.where(mask, uint16_seq, -100) stays uint16 under NEP 50 and wraps
-100 to 65436; cast the sequence to int32 first.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Tool-call inference
- chronos.serving.tool_calls: streaming <tool_call> parser that never
  leaks a partial tag; invalid blocks stay content.
- /api/chat returns Ollama message.tool_calls; /v1/chat/completions
  returns OpenAI tool_calls (string arguments, finish_reason
  "tool_calls"), streamed or not. Requests without tools are unchanged.

Multimodal preference stages on MLX
- DPO/ORPO: chosen and rejected condition on the prompt's media.
- GRPO: rollouts condition on the prompt's media (encoded once, same
  logits); the policy update re-encodes so encoders train.
- Distill: media reach the student; a text-only teacher is accepted.
- Data: MultimodalStreamingDPODataset + preference_collate,
  load_grpo_prompt_records, shared record_media resolver.
- Workbench: media data in DPO/ORPO/GRPO selects MLX under auto; an
  explicit PyTorch backend raises instead of dropping the media.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

- chronos/model/kernel.py replaces the minimind kernel: no clone, no
  sys.path changes, offline import; logits bit-identical on the existing
  checkpoints; one shared FP32 RoPE table (3.50 -> 0.50 MiB at 50M).
- Dense GQA mode works on Torch and MLX (both crashed before); BF16
  release checkpoints keep their dtype; trainer logs reach the Web UI.
- decode_visible keeps <tool_call>/<tools> tags that a custom tokenizer
  registers as special tokens (serving, workbench, GRPO rewards).
- MLX: compiled micro step by default (1.08x SFT, 1.18x pretrain measured
  on the 50M en->zh config) and activation recomputation with replayed
  draws and per-dtype packed region outputs (mlx 0.32.4 drops gradients
  of repeated region outputs; peak 5.62 -> 3.50 GB SFT).
- RoPE tables built single-threaded (1-ulp prefix mismatch at 18 threads).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

models/chronos-enzh-50m-sft/ packages the 50M Chronos v3 model trained
for 69 h on an Apple M5 Max: config, own 16k tokenizer, the trained
prompt, decoding settings, a driver (translate.py) and the evaluation.
Weights ship as a release asset (enzh-50m-sft-v1), not in git.

- Boundary: English -> Chinese only, one sentence of <= 128 tokens per
  call. zh->en is disabled because en->zh-only SFT erased it (FLORES+
  BLEU 0.00 after SFT vs 17.15 for the pretraining checkpoint).
- Measured (MLX bf16, greedy): FLORES+ devtest BLEU 33.48 / chrF 29.52,
  WMT22 32.60 / 28.83, WMT24++ 26.89 / 24.43 (sentence split).
- The driver checks the weights' SHA-256, splits sentences, refuses
  input outside the boundary and cuts repetition loops.
- NOTICE: weights are for research use only (corpus licences unknown).
- tests/test_model_enzh_50m_sft.py covers the boundary checks without
  loading weights; the sdist prunes models/.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

- download() uses certifi's CA bundle when installed. python.org macOS
  builds ship no system CAs, so urlretrieve failed with
  CERTIFICATE_VERIFY_FAILED. A failed download now removes the partial
  file and says how to fetch the weights by hand.
- The CLI runs the language check before building the Translator, so
  Chinese input is refused (exit 2) even when the weights are absent,
  instead of failing with "weights not found".
- Two tests cover both paths.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit c550537)

…oE, recompute policy, RoPE scaling, packing)

Blockwise causal attention (full-causal blocks use the kernel's causal mode),
chunked cross-entropy, chunked MoE dispatch, selective recompute policy on MLX
and Torch, training-time RoPE scaling (yarn/linear/ntk), document masking with a
separate lookahead document switch, a pretrain length curriculum and SFT packing.
All switches default to off. 64k full-model step: 24.5 GiB extra peak at 6.6k tok/s.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

…led document-reset fix

Adds recompute_scope=mlp, mla_upproj_recompute, lookahead_input_detach and
temporal_grad_span (all off by default), web UI fields in four languages, and a
real-corpus A/B report. Fixes reset_router_at_documents / document_masking crashing
inside the compiled LongRun step (found by training on the real corpus) and the
"false" string read as true for reset_lookahead_at_documents.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

… compile trace

- cluster_locality_mode="segment" now traces under mx.compile: document ids come from
  segment_ids_mx(input_ids, chunk=window) and the loss uses static shapes (B*S segment
  slots, unused slots weigh 0). Matches the NumPy ids and the Torch loss.
- CompiledStep snapshots weights, RNG key and side outputs before a first trace and
  restores them when the trace fails, so the eager fallback no longer trains on tracers.
- tests/test_segment_compile.py (5 cases; 4 fail on 0.3.2a2).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

MLX does not bounds-check gathers, so an id >= vocab_size read neighbouring memory and
crashed the process at random. run_mlx_stage (every batch, GRPO prompts) and LongRun now
raise a ValueError naming the id range.

test_mlx_training_six_stage_loss_paths crashed about one run in two on CPU, on clean dev
as well: its GRPO stage tokenizes with the bundled 6400-token tokenizer while the model had
vocab_size=64 (ids up to 6170). The test now uses vocab_size=6400 (25/25 runs pass).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

…task-neutral kernel

Loop (Torch and MLX): recurrent_order "block", recurrent_later_topk, recurrent_later_groups,
lambda_iteration_cluster, lookahead_teacher_scope, recurrent_dt_param "sigmoid", and
recommended_dt_init(depth) = min(0.5, 2 / depth) (fixed integration horizon T = depth * dt).
Kernel defaults are unchanged. The planner now recommends block order, sigmoid and the horizon
init for a looped plan; presets_v3.json is regenerated.

Fixes: the MLX clip gradient on dt_min/dt_max is now the Torch clamp gradient; the MLX depth
jitter is drawn once per optimizer step under gradient accumulation.

Auto-tune: the probe trains on real dataset tokens, and recurrent_order and recurrent_dt_init
join the search space. Web UI: new fields and four-language help.

Tokenizer: chronos/assets/tokenizers/general_32k (32768 tokens, task-neutral control ids,
licences in NOTICE) is the planner default for new plans.

Kernel text no longer names a translation task; the translation sections moved to
demos/translation. CHANGELOG records the physical mechanism, the measurements (one seed per
arm, 15M parameters) and the known limits.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

…-entropy, demo model under demos/

Planner: a looped plan also sets recurrent_later_topk=0, recurrent_later_groups=0,
lambda_iteration_cluster=0, lookahead_teacher_scope="first", recurrent_dt_min=0.05 and
recurrent_dt_max=1.0, each with a sources entry stating the 15M measurement behind it. The Web UI
shows them as planner recommendations; presets_v3.json is regenerated.

Auto-tune: all nine loop knobs are searchable (later_groups 0 or 2, never 1; block order only).
The loop probe has 4 block layers, clamps a drawn dt_init into the drawn range and logs each
trial's median step time. The probe now trains on chronos_loss_term and scores cross-entropy:
the old score added the weighted regularizers, and lambda_lookahead never reached the probe.

Web UI: Field.as_dict dropped 0 and 0.0 (0 == False), so min=0 and recommended=0 never reached the
page; it now drops only unset values. Help text gives the measured reason in four languages.

Translation demo: models/chronos-enzh-50m-sft and its driver test moved under demos/translation;
CI runs the driver test in its own step. The main READMEs name the demo in one pointer paragraph;
the measurement conditions moved to the demo READMEs. A guard test checks both.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Kernel code is unchanged from 0.3.3a1. This release ships the final
documentation and figure set:

- docs/KERNEL_ARCHITECTURE.md is the top-level kernel walk-through,
  with four code-generated architecture figures under
  docs/assets/arch/ (reproducible generator + symbol config).
- Per-component chapters MOE_ROUTING, CLUSTER_LOCALITY, RESIDENCY,
  LAZY_EXPERTS, LOOKAHEAD, LOOP_DEPTH, LONG_CONTEXT, SPEC_DECODE,
  each with source-line references and measured numbers traced to
  recorded JSONs under experiments/figures/source_data/.
- 18 redesigned data figures with editable and path-outlined SVGs
  under docs/assets/; a verify_values.py numeric audit (87 checks /
  880 numeric entries; exact except a 1e-12 tolerance on the
  sigmoid inverse) under experiments/figures/.
- Evidence ledger docs/REPRODUCIBILITY.md, symbol table
  docs/SYMBOLS.md, Chinese reading guide docs/READING_GUIDE_ZH.md.
- docs/README.md rewritten with a reading-route diagram, a kernel-
  interior gallery, and an at-a-glance measured-evidence row.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Resolves MANIFEST.in by keeping release-0.3.3's expanded prune list
(Translation Training Data, demos, chronos/mlx/training/recipes).

Resolves README.md / README_zh.md by removing the duplicate
'Released model: chronos-enzh-50m-sft' section that came from main
and pointed at the old `models/chronos-enzh-50m-sft/` path.
release-0.3.3 moved the model demo to `demos/translation/` and the
docs in this branch reference it there throughout, so the merged-in
section was both duplicated and pointing at a path that no longer
exists in this branch.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…test

tests/test_lazy_probe.py imports fastapi.testclient inline for the
webui-side assertion. The main CI workflow (ci.yml) already installs
fastapi/uvicorn/httpx alongside the base deps; the legacy
python-app.yml workflow did not, so its build job failed with
ModuleNotFoundError: No module named 'fastapi'. Mirror the ci.yml
install line.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>