JVNAUTOSCI-2596: freeze disposable read-only A1 candidate

[witbrock]

User outcome

Provide one genuinely thin, permissive comparison path for the authorised six-run A1 experiment, so evidence rather than another prewritten controller theory can determine whether A2 or any deletion slice should exist.

Material changes

• Adds one standalone, non-production script using the OpenAI Responses SDK directly and existing MCP read implementations.
• Freezes one concise intention prompt, gpt-5.6-luna, stateless/store-false Responses calls, two tool batches, Michael's actor scope, and a broad 52-tool read-only palette.
• Rechecks only the hard read-only boundary before advertisement and invocation.
• Emits one plain local JSON transcript; transcripts may contain private material and must not be committed.
• Adds four focused tests for stateless reasoning/tool continuation, malformed provider correlation, and the read/write boundary.

This PR is the branch-visible freeze artefact for JVNAUTOSCI-2596. Candidate base: 95c6a4caad1f50b0845f2e654b97522cc586426b. Frozen candidate commit: 207178f0d4a7133a70aa320cd267e311797f7407.

The earlier pre-freeze SHA was superseded before any case was supplied or seen solely to remove import-time path mutation and avoid rebuilding a constant set on each tool invocation. Prompt, model, tool palette, actor, limits, and experiment protocol did not change.

Evidence

• python -m pytest -q tests/test_thin_readonly_turn.py — 4 passed.
• Ruff check and format check passed for the two candidate files.
• Python compilation passed for both files.
• Read-only preflight against the configured Atlas-backed catalogue passed: all 52 frozen tools were registered as read.
• No evaluation case has been supplied, seen, or run.

This evidence supports candidate mechanics and the read-only boundary only. It says nothing yet about usefulness or relative performance.

Decision boundary

Do not merge this PR. It is a disposable A1 candidate, not a proposed production controller. Source, prompt, model, tools, actor, and limits are frozen before the cases are revealed.

Michael supplies three standalone cases after the current-Von comparison profile is also frozen. Each case is attempted once here and once on current Von: exactly six submissions, no retry, rephrase, tuning, replacement, or seventh run. H1/JVNAUTOSCI-2597 belongs to Michael and alone decides CONTINUE, REVISE, REORDER, or STOP. A2 and the first deletion slice have no authority until that decision.

[Copilot]

Pull request overview

Adds a disposable, non-production “thin read-only turn” A1 candidate for JVNAUTOSCI-2596, intended to exercise a frozen OpenAI Responses-based loop over an existing read-only MCP tool palette and emit a local JSON transcript for comparison runs.

Changes:

• Introduces scripts/try_thin_readonly_turn.py, a standalone stateless Responses loop with a hard read-only tool boundary enforced via Internal MCP method definitions.
• Adds tests/test_thin_readonly_turn.py covering tool-continuation mechanics, malformed provider tool-call correlation handling, and read/write boundary enforcement.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
scripts/try_thin_readonly_turn.py	Implements the disposable stateless Responses-based candidate loop and read-only tool boundary checks.
tests/test_thin_readonly_turn.py	Adds focused tests validating tool continuation, provider protocol errors, and boundary rejection behaviour.

Comments suppressed due to low confidence (1)

scripts/try_thin_readonly_turn.py:348

• After moving the sys.path manipulation behind a helper, call it early in main() (before importing from src...) so the script still works when executed from outside the repo root, while keeping imports side-effect free for test imports.

    args = parser.parse_args(argv)

    try:

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated no new comments.