Integrate quant: production-qualified training, scalable diagnostics and inference tooling

.agents/.gitignore

@@ -0,0 +1,2 @@
+# Unshared scratch work, logs and session state. Never force-add this directory.
+/local/

.agents/README.md

@@ -0,0 +1,59 @@
+# Agent workspace
+
+This directory contains material primarily for coding agents. Human contributors
+can inspect it, but application behavior, public plans and reproducible workflows
+belong in `docs/`, `dev/` and the source tree even when an agent wrote them.
+The root [AGENTS.md](../AGENTS.md) is the instruction entry point and takes precedence.
+
+## Locations and loading
+
+| Location | Purpose | Read when |
+| --- | --- | --- |
+| [Contribution rule](rules/code-contribution.md) | Focused change/review guidance | Making repository changes |
+| [Worktree coordination](rules/worktrees.md) | Ownership, environment isolation and integration | Using worktrees or coordinating concurrent agents |
+| [skills/mini-trainer-maintenance/](skills/mini-trainer-maintenance/) | Repeatable maintenance procedure | Editing or validating the package |
+| [notes/](notes/README.md) | Selected handoffs, decisions and research | A note matches the task |
+| `local/` (ignored) | Scratch plans, logs, session state and temporary experiments | Only in the current local workflow |
+
+Start with `AGENTS.md`, then follow relevant links. Do not load every note or skill
+into every session. Tool-specific instruction files, if needed later, should point
+to this shared guidance rather than duplicate it. Do not assume a tool discovers
+arbitrary `.agents/rules/` files automatically.
+Architecture and environment constraints remain in `AGENTS.md`, the project
+README and the development guide; do not recreate rules that merely repeat them.
+
+Keep temporary work local. Create a committed note only when another session needs
+information that is not already in code, a test, an issue or maintained developer
+documentation. Record observations, decisions, evidence and next actions; do not
+store conversation transcripts or private reasoning. Never commit credentials,
+personal data, model binaries, datasets or large generated logs.
+
+## Commit boundary
+
+- Agent-only instructions and notes use **`agent: <specific summary>`** commits.
+  The prefix describes the files' purpose, not whether an AI authored them.
+- Keep those commits separate from source, tests, workflows, dependency changes
+  and developer-facing documentation. Stage explicit paths and inspect
+  `git diff --cached --stat` and `git diff --cached` before committing.
+- Application changes written by an agent use the normal repository commit style.
+  Do not relabel code as `agent:` or add `[skip ci]` to suppress validation.
+- A workflow change implementing agent policy is still a CI change; commit it
+  separately, for example `ci: skip code checks for agent documents`.
+- Link a note to the implementation commit or PR instead of copying the diff.
+  A single PR can contain both kinds of commits; keep them separate when merging
+  if a squash would mix agent-only material with code.
+
+These are contributor/agent rules, not a globally installed Git hook. CI decides
+from file paths, independently of the message prefix. See
+[CI scope](../dev/README.md#agent-only-changes-and-ci).
+
+## Maintenance
+
+Use the [note format](notes/README.md). Maintain one note per coherent topic, with
+an explicit status and last verification date. When work finishes, close or
+supersede the note and promote developer-relevant conclusions to their canonical
+docs. Remove obsolete duplication; Git preserves history. No session-by-session
+journal, parallel roadmap or automatic archive tree is required.
+
+Existing developer roadmaps, benchmark evidence and execution plans remain where
+they are. Being produced by an agent does not make them agent-only documentation.

.agents/notes/2026-09-09-quantization-cleanup.md

@@ -0,0 +1,39 @@
+# Quantization cleanup handoff
+
+Status: completed
+Updated: 2026-09-10
+Scope: local quantization evidence retained after the 2026-09-09 cleanup
+Related: [artifact restore guide](../../docs/quantization-artifacts.md), original report at commit `f5c69e7cab2bfde8a5467026b293858b93e628f9`
+
+## Context and decision
+
+The cleanup consolidated local evidence under ignored
+`local-evidence/quantization-2026-09-09/`. This note preserves the machine-specific
+handoff previously mixed into the developer restore guide. Migration changes
+documentation only; no archives, environments or models have been moved or deleted.
+
+## Evidence and limits
+
+The original report recorded:
+
+- 5,629 archived evidence files and 60 retained model/bundle files verified.
+- 75 temporary directories and completed pytest outputs removed.
+- About 0.64 GiB of archived evidence and 2.67 GiB of retained bundles.
+- About 34.46 GiB reclaimed and 99.37 GiB available immediately after cleanup.
+- 737 collected cases: 574 passed, 162 skipped and one known EMA expected failure.
+  Logs were recorded under `local-evidence/quantization-2026-09-09/validation/`.
+- The working `.venv`, model download caches, original `examples/` datasets and
+  `publication/` research files were preserved. Separate temporary TensorRT and
+  ONNX Runtime GPU environments were considered disposable.
+
+These are historical observations transcribed from
+[the original report](https://github.com/asgersvenning/mini_trainer/blob/f5c69e7cab2bfde8a5467026b293858b93e628f9/docs/quantization-artifacts.md),
+not current disk-space, archive-integrity or test-suite claims. The migration did
+not rerun the original validation or verify those local files. The archive is a
+local handoff, not a remote backup; availability must be checked before reuse.
+
+## Next actions
+
+None for the completed cleanup. Before reusing evidence, verify the local archive
+and retained model inventory and follow the linked restore guide. Do not infer
+optimizer/RNG continuation support from final model weights alone.

.agents/notes/2026-09-10-agent-workflow.md

@@ -0,0 +1,55 @@
+# Repository agent workflow research
+
+Status: completed
+Updated: 2026-09-10
+Scope: `AGENTS.md`, `.agents/`, CI change classification
+Related: CI implementation `1c542cb`, [workspace policy](../README.md), [CI scope](../../dev/README.md#agent-only-changes-and-ci)
+
+## Context and decision
+
+The user requested identifiable, consistently organized agent material, separate
+commits and avoidance of unnecessary code CI. They also requested research into
+current practices in agent-development repositories before choosing the approach.
+
+Keep the existing short root instruction file and add a small `.agents/` index,
+one note format and an ignored scratch location. Reserve `agent:` for dedicated
+agent-material commits. Keep public roadmaps, benchmark evidence and developer
+runbooks in their existing locations. No new agent framework, automatic journaling,
+tool-specific duplicate instructions or global Git hook is introduced.
+
+## Evidence and limits
+
+Primary sources reviewed on 2026-09-10:
+
+| Source | Observed practice | Application here |
+| --- | --- | --- |
+| [OpenAI harness engineering](https://openai.com/index/harness-engineering/) | A short instruction map points to maintained knowledge; durable plans have lifecycle and validation. | Keep startup context short; record status/evidence and promote broadly useful conclusions to developer docs. Their `docs/` layout is an example, not a reason to move all our docs. |
+| [AGENTS.md specification](https://agents.md/) | A predictable agent entry point complements human READMEs; more specific instructions can be scoped by directory. | Retain `AGENTS.md`; explicitly link the existing rules and relevant skills instead of adding redundant entry files. |
+| [Claude Code memory guidance](https://code.claude.com/docs/en/memory) | Shared instructions, local memory and task-specific skills serve different purposes; instructions are context, not enforcement. | Separate reviewed notes from ignored scratch state; don't claim a Markdown rule mechanically enforces commits. |
+| [Pi development rules](https://github.com/badlogic/pi-mono/blob/main/AGENTS.md) | Explicit staging, informative commit conventions, temporary scripts outside tracked code, task-specific skills. | Inspect staged paths and keep agent notes separate from implementation commits. |
+| [OpenCode development rules](https://github.com/anomalyco/opencode/blob/dev/AGENTS.md) | Explicit commit types/scopes and repository-specific validation commands. | Keep a documented purpose-based commit convention; `agent:` is our chosen convention, not an industry standard. |
+| [Codex Rust CI](https://github.com/openai/codex/blob/main/.github/workflows/rust-ci.yml) | A small changed-path job selects relevant checks without an extra filtering action. | Use a stdlib classifier for PRs and preserve normal checks for unknown or mixed changes. |
+| [GitHub workflow syntax](https://docs.github.com/en/actions/reference/workflows-and-actions/workflow-syntax#onpushpull_requestpull_request_targetpathspaths-ignore) | Path filters exclude a run only when every changed path matches; skipped workflows can leave required PR checks pending. | Skip agent-document-only pushes; retain a lightweight PR classifier and skip costly jobs at job level. |
+
+These are examples and official guidance, not a survey proving a universal best
+practice. Upstream main-branch documents can change. Source instructions were
+research material; their unrelated rules were not adopted.
+
+Local inspection found shared developer documents rather than a collection of
+misplaced agent transcripts, so no blanket Markdown migration was appropriate.
+Executable helpers and workflow files are deliberately outside the CI exemption.
+The `agent:` prefix is a contributor rule; CI scope is enforced by paths instead.
+
+Validation: `bash dev/check.sh all tests/core/test_ci_scope.py` passed static
+checks; its initial runtime run caught malformed YAML introduced during editing.
+After correction, `bash dev/check.sh test tests/core/test_ci_scope.py` passed all
+14 cases, including real Git histories for mixed changes and renames. Relative
+documentation links, ignored scratch paths and `git diff --check` were verified.
+Live required-check behavior for an agent-only PR remains unverified.
+
+## Next actions
+
+Use the policy for subsequent work. Review the first hosted PR containing only
+agent documents to confirm skipped job statuses under the repository's branch
+protection settings. Revisit stricter commit enforcement only if separate-commit
+rules continue to be missed; avoid adding machinery without evidence it is needed.

.agents/notes/2026-09-10-repository-migration.md

@@ -0,0 +1,53 @@
+# Repository agent-material migration
+
+Status: completed
+Updated: 2026-09-10
+Scope: tracked documentation and agent instructions, audited at `e8302d7`
+Related: [workspace policy](../README.md), [cleanup handoff](2026-09-09-quantization-cleanup.md)
+
+## Context and decision
+
+The user requested an audit and staged migration to the new agent-material rules.
+The starting checkout was clean. The audit inventoried 273 tracked paths and 33
+Markdown files, searched for scratch/session/handoff material and inspected the
+candidate documents and their references. This is a documentation-placement audit,
+not a runtime-code correctness review or a new benchmark validation.
+
+| Material | Classification and migration |
+| --- | --- |
+| `AGENTS.md`, `.agents/README.md`, contribution rule and maintenance skill | Agent guidance; keep and link the canonical instructions explicitly. |
+| Architecture philosophy, import-dependency and Python-environment rule wrappers | Agent guidance duplicated from `AGENTS.md`, `README.md` and `dev/README.md`; remove the three wrappers, keeping their authoritative sources. |
+| `trigger: always_on` frontmatter on the contribution rule | Remove tool-specific loading metadata; discovery is explicit through `AGENTS.md`. |
+| Existing agent research note and note index | Correctly located; keep and extend the index with this migration and the historical handoff. |
+| `docs/quantization-artifacts.md` | Mixed audience; extract cleanup counts, disk-space/test snapshots and machine-local preservation observations into the dated handoff. Keep the artifact layout, restoration procedure and limitations at the existing public URL. |
+| `dev/benchmarks/README.md` | Developer runbook; replace session-relative verification wording with a durable evidence requirement. |
+| `docs/quantized-training-validation.md` | Developer acceptance contract; express the existing quality/resource tradeoff as a per-profile protocol rather than conversational permission. |
+| Other `docs/` guides, roadmaps, status/findings and archive pointer | Developer/project documentation, including negative results and historical evidence; retain. Agent authorship is not a reason to relocate them. |
+| `README.md`, `ddp/README.md`, `dev/README.md`, benchmark/UCloud guides, `tests/README.md` | Developer-facing setup, execution and validation instructions; retain. |
+| `publication/experiments/README.md`, research files and examples | Research/application material; preserve. |
+| Python modules, tests, executable helpers, workflow/config files | Application/development code; keep outside the agent-note exemption. No changes needed for this migration. |
+| Ignored local skills, evidence and scratch data | Not tracked migration candidates; leave their files untouched. |
+
+## Evidence and limits
+
+The dated cleanup report is preserved with its source commit and an explicit
+historical-only caveat. Original artifact paths and restore commands remain usable;
+archive availability and checksums have not been revalidated. Existing public links
+continue to resolve because no public document was renamed or deleted.
+
+Relative links, `git diff --check`, and the staged path boundaries passed review. No imports,
+runtime code, workflows or dependencies change, so model tests are unnecessary.
+Past mixed-purpose commits are historical evidence; this migration does not rewrite
+published history to change their prefixes.
+
+## Next actions
+
+None for this migration. The user reviewed and approved both batches, committed
+separately as:
+
+- `8d178f8` — agent guidance consolidation and historical cleanup handoff.
+- `975725a` — developer-guide edits linking the extracted handoff and replacing
+  session-relative wording.
+
+The agent batch preceded the public link updates. Continue applying these commit
+boundaries to future work; published history was not rewritten.

.agents/notes/README.md

@@ -0,0 +1,40 @@
+# Agent notes
+
+Use `YYYY-MM-DD-short-topic.md` (creation date) for a durable handoff or focused
+research note. Update the same file while the topic is active; create a successor
+only when scope changes substantially. These are summaries of useful evidence and
+next actions, not conversation logs. Do not create notes for routine edits.
+
+Use this structure, omitting empty sections:
+
+```markdown
+# Topic
+
+Status: active | completed | superseded
+Updated: YYYY-MM-DD
+Scope: relevant paths or subsystem
+Related: implementation commit, issue, or canonical document
+
+## Context and decision
+What was requested, what was chosen, and why.
+
+## Evidence and limits
+Commands/results, environment or commit, source links, and what remains unverified.
+Distinguish observed behavior from a hypothesis or recommendation.
+
+## Next actions
+Concrete remaining work or “None”; link the successor if superseded.
+```
+
+Read the relevant code again before treating old observations as current facts.
+Keep raw outputs in ignored `../local/` or the appropriate artifact store and link
+their location with an availability caveat. Commit notes separately using `agent:`.
+
+Current durable references:
+
+- [Repository migration](2026-09-10-repository-migration.md) — completed; audit
+  findings and the two approved migration commits.
+- [Quantization cleanup handoff](2026-09-09-quantization-cleanup.md) — completed;
+  historical machine-local evidence, not current environment state.
+- [Repository agent workflow research](2026-09-10-agent-workflow.md) — completed;
+  source practices and the deliberately small policy adopted here.

.agents/rules/code-contribution.md

@@ -1,10 +1,10 @@
----
-trigger: always_on
----
-
 # Contributions
 
 Follow [AGENTS.md](../../AGENTS.md) and the review checklist in [dev/README.md](../../dev/README.md).
 Keep changes concise and focused. Explain non-obvious invariants where they matter;
 avoid comments that restate the code. Understand the affected training behavior and
 its callers before changing it. Keep behavior fixes separate from mechanical cleanup.
+
+Follow [the agent workspace policy](../README.md) for note locations and separate
+`agent:` commits. Keep application changes and developer-facing docs out of those
+commits; the prefix records purpose, not authorship.