[codex] simplify provider-neutral negative knowledge module

[hch-wang]

What changed

• replace the multi-file package and provider-specific runtime with a single zero-dependency negative_knowledge.py module
• expose four small functions: curate, validate, append, and load
• add strict record validation, evidence truncation, JSONL persistence, a concise README, an offline example, MIT licensing, tests, and CI
• route optional fresh reproduction runs through the provider-neutral NK_AGENT_COMMAND JSON protocol
• anonymize provider/model labels in the current tree as primary, secondary, or provider-redacted, including archived filenames

Why

The reusable surface should be small enough to copy directly into another project and should not require a particular model SDK. The previous implementation coupled curation, file tools, prompts, CLI behavior, and one provider into a larger package.

Impact

This is a breaking API simplification: callers should replace NKCurator usage with curate(...) and use append / load for JSONL memory. Current-tree model provenance is anonymized; Git history is unchanged.

Validation

• python3 -m unittest discover -s tests — 8/8 passed
• Python 3.9 syntax parse passed
• offline quickstart passed
• clean wheel build and no-dependency install passed; pip check passed
• §3 claim verification: 31/31
• §4 claim verification: 20/20
• appendix claim verification: 54/54
• Table 1 token medians reproduced exactly
• full working-tree search found no provider/model brand terms
• git diff --check passed

[hch-wang]

Notes after reviewing this branch and running the suites locally.

The core direction is good: collapsing to a single dependency-free module and making curation provider-neutral (the caller supplies backend) is a clean design, and the tests / CI / richer CITATION.cff are welcome. A few things I'd want resolved before merging:

1. Branches before the release polish → conflicts with main.
This is based on f915d96, before main gained its LICENSE, README badges, arXiv links and packaging metadata (f8fe97a). It currently shows as conflicting, and the rewritten README.md drops the badges, the arXiv link, the overview figure, the schema section and the BibTeX — regressing the release-ready front matter instead of building on it.

2. The new validate() rejects the repo's own records.
It is stricter than the data this project actually produced:

• length caps (200 / 300) sit below the observed maxima in reproduction/ (attempted_route 332, observation 300, rationale 404, recommended_alternative 402);
• real records carry extension fields (round, relationship_to_round1, is_trivial, …) that are flagged as unknown field;
• LAYERS is missing values that appear in the artifacts (hypothesis_failure ×14, measurement_failure ×3).

Net effect: only 11 of 67 bundled per-attempt NK records pass validate(), so load() / curate() would raise on legitimate records. Deep (cross-round) records aren't representable at all under the six-field schema.

3. Anonymizing the reproduction/ artifacts.
The sonnet → primary / haiku → secondary scrub (most of the ~400-file diff) lowers transparency for a paper-reproduction repo and makes it inconsistent with the paper, while not actually removing the names (git history is unchanged, as the description notes). Better to keep the module provider-neutral but leave the reproduction artifacts as the honest record of what was run.

4. Minor. reproduction/agent_command.py calls shlex.split() in POSIX mode, which mangles Windows paths (WinError 2); 1 of the 8 tests fails on Windows as a result.

---

Reconciled branch. I pushed cleanup/provider-neutral-module that keeps the good parts of this PR and addresses the above:

• single-file zero-dependency module + tests + CI + the richer CITATION, rebased on main (no conflicts);
• merged README — badges / arXiv link / overview figure / BibTeX kept, and the new curate / validate / append / load API documented;
• relaxed length caps (→ 600 / 1200), extension fields allowed, LAYERS widened to the taxonomy actually used → all 67 bundled records validate;
• shlex fixed for Windows → 8 / 8 tests pass;
• reproduction artifacts left un-anonymized.

Locally verified: unittest 8/8, offline quickstart OK, §4 20/20, appendix 54/54.