research: publish recursive feedback protocol and harness

[Steake]

Summary

This PR publishes a focused, falsifiable research package for GödelOS output-to-input recursion:

• reconstruction manuscript: The Observer Is in the Loop
• registered five-condition protocol
• claims and provenance ledger
• frozen 24-prompt bank with SHA-256
• standard-library experiment runner for OpenAI-compatible endpoints
• exact JSONL trace schema, per-run hash chain, and manifest
• dependency-light lexical and cycle analyser
• 10 offline contract tests using deterministic replay
• README and whitepaper-index corrections that distinguish initialized components from per-request activity

Evidence status

This is deliberately a reconstruction manuscript and registered-report draft, not a results paper.

The repository audit did not locate the original raw traces, immutable model revision, complete prompt bank, decoding configuration, or run manifest for the early experiment. The PR does not invent those details or report remembered trajectories as evidence.

The design treats observer-state text as a causal intervention. It compares accurate state, matched-format sham state, and no-state conditions.

The architecture audit also separates three claims:

1. the cognitive pipeline initializes 23 named components;
2. initialization does not show that all 23 execute or causally contribute on every request;
3. no general task selector across all 23 was found in the audited code, although narrower routing exists.

The seven global-workspace state sources are documented as a separate taxonomy.

Verification

• python -m unittest discover -s tests -p 'test_*.py' -v (10 tests passed)
• python -m compileall -q experiments tests
• all experiment JSON files parse with python -m json.tool
• sha256sum -c prompt_bank.v1.sha256 (verified)

Formatting tools named in the repository guide were not installed in the execution environment, so Black, isort, and mypy were not run. Added Python files were checked for 88-character line length.

Review focus

1. Are the five conditions sufficient to separate exact recursion, persistent history, and observer-state effects?
2. Is the claims ledger fair to the current implementation?
3. Are the trace fields and hash chain adequate for a first public run?
4. Should the first confirmatory collection use the full 24,000-call matrix or a separately labelled variance pilot first?

Context: #136. The broader canonical-whitepaper tracker remains #86.

CI status

GitHub created both matrix jobs but did not start either one. The job annotation states: “The job was not started because your account is locked due to a billing issue.” No checkout, dependency-install, test, report, or artifact step ran.

The local verification above is therefore the only executed validation for this head until the account issue is cleared and CI is rerun.