TopoLab

Imported from keith1117/TopoLab (AGENTS.md). Install upstream with npx skills add keith1117/TopoLab. Copyright stays with the author.

TopoLab agent instructions

Read before starting work

At session start, slice start and context recovery, read these files in order:

1. PROVENANCE.md
2. docs/numerical_conventions.md
3. docs/reference_baseline.md
4. docs/planning/TopoLab_reimplementation_strategy.md
5. docs/planning/TopoLab_development_timeline_and_resources.md
6. docs/planning/TopoLab_v2_delivery_reassessment.md (English working reference)
7. docs/planning/TopoLab_v2_development_roadmap.md
8. docs/project_status.md
9. The contracts and latest validation reports relevant to the requested slice.
10. docs/repository_policy.md before staging or publishing changes.

Current state and execution entry

Historical v1 is a usable interim delivery. The owner retains FULL_V2_OPEN_NO_ADMITTED_ROUTE, fixed P/17/unrepaired,4GiB and independent preflight. Uniform stays default; full v2 still requires reproducible same-quality learned end-to-end acceleration. I1 public software is published through PR197/198, main57c73b8 and all four main CI contexts passed. Its mandatory selected-host pre-entry lifetime UNKNOWN closes NO_ADMISSIBLE_ROUTE, not measured>4GiB or scientific acceptance. Read current status for evidence, retained failures/costs/quotas/seals and the one next item.

Read the R1 plan after the English reassessment. Reassessment Section4 methods are effective; the roadmap Section6 separates current rules, historical observations and pending acceptance advice. The English roadmap controls scientific Gate order; README is the public entry. No scientific threshold, budget, access or concrete environment admission changes. B4.51 FAIL_PROCESS_SOURCE_FREEZE/consumed1/1 quotas, D1 STOP/unlaunched0/1/FULL60, owner-closed rescue and solver-internal rejection remain; B4.52 NOT_SELECTED_BY_R1. Known-paid12721.16417100014444982>7200 cannot pass unchanged all-paid feasibility; full fit PENDING4GiB and every old UNKNOWN/seal remain. Read applicable frozen contracts before dependent work; unresolved substantive conflicts stop that work.

B guest-capacity adoption remains ADOPTED_PROSPECTIVELY; ENVIRONMENT_NOT_ADMITTED, for compatible future guest contracts only; excluded host overhead UNKNOWN. The owner separately adopted B4.30-style RSS/full-exit empirical acceptance for NEW M2 native identity only, retaining4GiB and independent preflight. This is not a-priori allocation assurance or a change to old I1/D1/B contracts.

M2 H stage A is closed through PR202/203/main65bbbd3. Old R/v1 remains STOP_NO_RETRY / FAIL_FULL_NAMED_WALL_CHARGE_BINDING:109.224s+10>immutable20s, final serializer exit1, full RSS UNKNOWN, old reads0, R0/R1 closed unlaunched. Failed FULL120+shared FULL60=180/1860s; closed1680s never transfers. Source PR204/ main43ffed0 preceded entry; result PR205/main3459dd5 and all four main CI passed.

The owner expressly authorized NEW R/v2 under the protocol: /root3h and new different-author /root/review_r_v22h; six once-only FULL charges120/120/120/780/660/60=1860s and exact40-metadata/12-checkpoint/528-train/ 18-uniform original-byte/temporary-parse scope. Preserve v1 frozen bytes/STOP/costs. Reviewed source passed the five mandatory normal-permission checks (2203 passed/1 platform skip) and PR206/latest-head four CI jobs, then merged as main04d14c9. The earlier default-sandbox7 native permission failures remain recorded. Exact main CI39208671873 quality was still running at06:22:14 UTC. The independent window ends06:34:17 UTC, so later exporter-r1 FULL780 latest start06:21:17 had passed before source admission could be frozen. Close this finite execution unlaunched: no R/v2 plan/root, named role, old source stat/read or restricted I/O charge; source/resource acceptance remains UNKNOWN_NOT_RUN, not formal FAIL. That exact main CI later finished with all four jobs successful; it does not reopen the elapsed window. Publish the zero-launch result normally. Do not use the old grant again, start a partial prefix, B/C, old backup/seal, expanded input, retry/reset or an automatic successor. Retain4GiB, independent preflight, all old costs and gates.

Provenance boundary

The Hack3D reference repository has no declared open-source license. Do not copy, translate, patch, vendor, or redistribute its code, comments, file structure, or figures. Implement from published equations and independent derivation. Use the fixed upstream commit only for historical behavior comparison, and record any comparison or new source in PROVENANCE.md.

Development rules

• Specify deliverable-linked acceptance and obtainable resource evidence before choosing an algorithm; satisfy existing applicable contracts. Debug public software and complete checks before freeze; enforce preflight/freeze/STOP.
• Measure P/17 failures/quality gap, fully charged speed, resource admission and used/remaining budget; unmeasured stays UNKNOWN. At most two ineffective capability slices require a route decision, never automatic continuation.
• Before commit/push, recheck current rules/status and actual records; do not rely only on memory or rerun completed formal experiments.
• Keep each change scoped to one numerical behavior with tests.
• Freeze conventions in docs/numerical_conventions.md before relying on them.
• Prefer the current small module plan; split modules only when stable responsibilities or multiple callers justify it.
• Do not add FastAPI, React, Redis, PyTorch, data generation, or ML before gate N2.
• Do not claim validated, scalable, or accelerated until the documented evidence gate is complete.
• Never commit generated datasets, run artifacts, caches, secrets, or device IDs.
• Follow docs/repository_policy.md: keep protocols and validation reports in Git, keep generated evidence external, and stage only reviewed paths.
• Close slices by updating current status, one history entry and the relevant roadmap section/index. Keep this file concise; do not append the full chronology. Update README for public usage/claims changes or final delivery, not every slice.
• Name branches by deliverable without codex or other agent-name tokens; merge PRs only after CI passes and report the next slice after each slice.

Required validation

Run before every commit:

uv sync --dev --locked
uv run ruff check .
uv run mypy src
uv run pytest
git diff --check

If a numerical test requires a tolerance change, explain the scale/conditioning reason in the validation report; do not silently loosen it.