contrib: MolecularFullerene - dsh-schema-ablation

[MolecularFullerene]

Summary

This adds an independent controlled ablation of the DeepSeek Harness
first-request tool surface, plus a small second-request transport protocol pilot.
It is adjacent to, but does not replace or modify,
contributions/xiaobright-deepseek-v4-harness.

Main observations:

• With the official Minimal system, contexts=[], exactly two visible tools,
  and reasoningEffort=max, the persistent-bash schema bundle produced
  17/20 minimal-like lexical trajectories versus 6/20 for the one-shot
  bash bundle.
• The editor/read marginal comparison was 13/20 versus 10/20; the evidence for
  that file-tool main effect is weak.
• All 40 first actions were legal and task-relevant. We need / Let me is
  reported only as a trajectory diagnostic, not as an ability score or proof
  of two internal personas.
• A retain/drop reasoning_content × same/new session protocol pilot used
  one independent first-request source per treatment. All four cells (n=1 each)
  were accepted and returned the preregistered simple JSON answer. This does
  not establish equivalence on complex agent tasks.
• The evidence audit keeps the historical Project2 98/99 runs attached to
  their recorded legacy Windows pwsh/read → 25 tools configuration rather
  than the later exact-Minimal pair. The docs-only provenance request is
  tracked in xiaobright/dsh-anchored-standard#60.

Included material

• one contribution README with environment and limitations;
• three deidentified reports;
• three aggregate JSON summaries;
• append-only index rows in the two repository READMEs.

No raw trajectories, complete reasoning, request bodies, API/user/session IDs
or hashes, keys, absolute paths, Project2 prompt text, evaluator, or hidden
tests are included.

The reusable CLI, fixtures, and security tests live separately at
MolecularFullerene/deepseek-v4-pro-harness-ablation.

Validation

• python3 scripts/check_pr.py --base origin/main --head HEAD: 24/24 checks
  passed.
• git diff --check origin/main...HEAD: passed.
• The contribution is one commit on top of upstream
  aceaa0612ef7c9e7558f9dd5a6b694662cdb6e66.

Important limitations

• The shell factor changes a full schema bundle, not persistence semantics
  alone.
• The 40 trajectories use two prompt templates and are not 40 independent
  tasks.
• Prompt, identity, batch, and time are partially confounded; the reported
  Fisher p-values are nominal.
• The model is a mutable service alias with no exposed immutable server build.

This is opened as a draft so the maintainer can first review scope,
provenance, and wording.