config.example.json ships a `_comment` in `retrieval`, which sialib rejects: every fresh install lands in `state degraded`

[Enor75]

Summary

config.example.json ships a _comment key inside the retrieval block, but sialib.py accepts only associative_rerank there. Since install.sh installs that example file as the instance config, every fresh install lands in state degraded with a permanent config error.

Evidence

The validator is an exact set difference, with no _comment exemption:

# bin/sialib.py:153-158
retrieval = value.get("retrieval", {})
if not isinstance(retrieval, dict):
    _record_config_error("retrieval-must-be-object")
else:
    if set(retrieval) - {"associative_rerank"}:
        _record_config_error("retrieval-unknown-key")

The shipped example carries two keys:

$ python3 -c "import json;print(list(json.load(open('config.example.json'))['retrieval']))"
['_comment', 'associative_rerank']

And the installer publishes that file as the live config:

# install.sh:9354
atomic_install_file "$REPO/config.example.json" ...

Observed

Immediately after a successful install (SIA 1.7.8, Omarchy/Arch):

SIA · state degraded · pulse 50
  readiness READY — corpus, PGLite, and graph reconciled
  chains   pass (sia:pass)
  errors   {'sense_custom:0': [{'config': 'config.json', 'error': 'retrieval-unknown-key'}]}
  config   [{'config': 'config.json', 'error': 'retrieval-unknown-key'}]

Removing the single _comment key from retrieval and restarting sia-brainstem.service clears it:

SIA · state thinking · pulse 52
  readiness READY — corpus, PGLite, and graph reconciled

Note that readiness was READY and every chain passed in both cases — only the config validator dissents, so the degraded badge is misleading about the brain's actual health.

Impact

Every install starts degraded. The error names the key but not which block it came from, and the sense_custom:0 prefix points at custom senses, which is where I first looked — the offending key is in retrieval.

Suggested fix

Either accept _comment in retrieval the way the other blocks tolerate it, or drop the key from config.example.json. The first is more consistent: judge, senses, skills, custom_senses and chains all carry a _comment and validate fine.

Environment: SIA 1.7.8, Omarchy/Arch, Linux 7.2.3-arch1-3, Python 3.14.7.

[AnubisQuantumCipher]

Thanks for the precise report. I have reproduced the configuration-shape mismatch in the current source: the shipped retrieval._comment and the strict retrieval-key validator disagree, and the installer publishes that example as the initial live configuration. I am treating the resulting fresh-install degraded state as a real release blocker.

Triage is starting now on fix/perception-integrity: first a fresh-install regression that proves the shipped example is accepted through the same config loader used by the brainstem, then the smallest consistent schema correction plus the broader config/install checks. There is no released fix yet, and I will not describe the branch repair as available until a tested non-main checkpoint is published and a release containing it exists. main remains frozen for the existing Omarchy compatibility constraint.

[AnubisQuantumCipher]

Triage found that this exact defect was already repaired inside the broader integrity checkpoint 83c45c6. The repair admits retrieval._comment, validates that it is bounded text, and adds a regression that copies the shipped config.example.json into the live config path and loads it through sialib.load_config().

Current verification on fix/perception-integrity:

python3 -m unittest tests.test_retrieval_policy
Ran 10 tests in 0.703s
OK

That fix commit is an ancestor of the published, full-suite-tested source checkpoint 70d4814, available on fix/issues-7-8-verified. The sequential full suite at that checkpoint reported:

Ran 3497 tests in 5056.742s
OK (skipped=1)

Delivery status: this is a source checkpoint for testing, not a release. v1.7.8 still contains the defect, the issue stays open until a release contains the repair, and frozen main has not been changed.

[AnubisQuantumCipher]

Closing the implementation issue at the owner’s request; release delivery is now tracked in draft PR #15. Fix: 83c45c6 . Published candidate: 5cce3e1.

The shipped-example/load_config regression and bounded comment validation are documented above: tests.test_retrieval_policy reported 10 tests, OK. The candidate’s recorded full suite passed: 3556 tests, OK (skipped=1). These are retained results, not a new run today.

The repair is published source, NOT a release. Closing this implementation ticket does not mean existing installations receive the fix; release review and promotion remain open in #15. Main remains untouched.