[robertohluna]

Merged and deployed to the running local services on September 9, 2026 using the repository owner's explicit one-time review exception.
Code-owner review protection was restored and verified immediately after each merge; required CI and administrator enforcement remained enabled.
This exception was not an independent code-owner approval or independent red-team certification.
Post-merge application CI passed.

Current setup docs still advertised an authentication override that runtime configuration ignored, and the CLI could fall back to a different local store when the intended API rejected its health probe.
This change fixes those reproduced behaviors, closes authority-reference gaps, and reconciles the current operator documentation with actual runtime and security boundaries.

• Honor literal OPTIMAL_AUTH_REQUIRED=true/false; reject invalid overrides and preserve configured values when absent.
• Send the configured bearer credential on API probes; refuse local fallback on HTTP errors or an unreachable explicitly selected API/credential; return failure for API action errors.
• Scan current contracts and bare Markdown references, with mandatory current-authority scanning.
• Update current boot/setup/deployment/API/CLI/store/readiness/version guides, indexes, registry classifications, and findings 4-6 / fixtures A-G evidence and retest instructions.
• Extend CODEOWNERS to nested Markdown and retained evidence.
• Preserve six exact positive/negative fixtures with SHA-256 hashes, causal lessons, structured red-team records, additive supersession, and recurrence classification.
• Evaluate candidates using the prior main commit's unchanged validator and 26 tests inside a container with read-only source mounts, no network, and bounded resources, before running candidate code on the host.
• Add an opt-in fresh Codex review suite that retains prompts, schemas, tool transcripts, exact-source citations, answers, evaluator hashes, and candidate commit identity.

Validation for this update: 79 Python regressions passed locally, including 32 authority cases, 17 trusted-evidence cases, 23 fresh-agent evaluator cases, and seven wrapper cases.
The container-based control-plane workflow passed for ad1e0d898f7d08d2e42563219bd2199cdbfe6451.
Its retained report verifies the previous main commit 51b6f8643f729ec134d2e9e8fd749e848bf76aeb, all 26 unchanged prior tests, and both validators' outcomes on the six exact fixtures.
Application CI also passed on that exact head.
All four real fresh Codex scenarios passed on the same committed snapshot: current authority, historical poisoning, documentation permission weakening, and missing mandatory authority.
Twenty exact prompt/schema/transcript/answer/stderr artifacts were retained and their hashes verified.
The first run exposed a terminal-event parser defect; its failed report was preserved, the exact event became a regression fixture, and a complete new run passed without changing the expected authority or permission decisions.
The prior runtime changes passed 2,025 Elixir tests; the latest application workflow reran the complete suite.

Initial harness/ledger installation is explicitly reported as pending-independent-bootstrap and requires independent code-owner review.
Subsequent revisions use the approved baseline's evaluator; candidate workflow changes still require owner review.
A direct local Python invocation is not a malicious-code sandbox.
The fresh-agent suite evaluates read-only authority decisions; it does not prove authenticated runtime execution or universal prompt-injection resistance.
No live user stores are used by these tests.

The default unconfigured local CLI can still intentionally run trusted Mix tasks when no API responds.
The manifest checks declared ownership and reference structure, not arbitrary prose semantics; reported build identity is not cryptographic server attestation.
No independent red-team hard-close certification is claimed.
Normal independent review requirements are restored for subsequent changes.