docs: correct stale direct-NIM claims after contextual-orchestrator migration

[seonghobae]

Summary

Documentation-only accuracy fix, found while surveying this cycle's goal (route Noema/OpenCode/Strix through contextual-orchestrator's orchestrator/free, eliminate direct NIM communication). That survey confirmed OpenCode review, Strix, Noema-review, and the hourly PR-review-autofix worker are already fully routed through the vendored contextual-orchestrator gateway — no live direct-NIM HTTP call exists anywhere in this repo's workflows. But three documents still described the write-capable autofix worker's retired direct-NIM architecture (NVIDIA_API_KEY bound to a hard-coded mistralai/mistral-small-4-119b-2603) as current fact, contradicting the actual .github/workflows/pr-review-autofix.yml and its own contract test (test_scheduled_autofix_routes_through_contextual_orchestrator, which explicitly forbids https://integrate.api.nvidia.com/v1 and "enabled_providers": ["nvidia-nim"]). This is the exact "follow-up doc cleanup, not attempted this pass" gap already flagged in docs/product-technical-gap-baseline.md.

Changes

• docs/doctoring/hourly-nvidia-nim-autofix.md: added a "Status (2026-08-31 correction)" section pointing at ADR-0003; rewrote "Provider contract", "Credential boundary", and "Rollback" to describe the current gateway/sidecar architecture (five KV-registered provider secrets feeding auto-discovery, never one bound directly); corrected two present-tense claims in "Verification contract" and "Scheduling and activation". Historical TDD-evidence records (exact-head RED/GREEN commit SHAs) are left untouched as dated history, per this repo's "append a dated note, don't rewrite history" convention. Removed three now-orphaned APA references to NVIDIA-specific model docs whose claims no longer appear in the body.
• docs/automation/hourly-review-repair.md: corrected the caller-contract prose (was: "uses OpenCode with NVIDIA NIM"; "the model credential is scoped exclusively to the two OpenCode execution steps") and the operator/verification checklists to describe five gateway provider secrets scoped to the sidecar-provisioning step, not a single directly-bound NVIDIA credential.
• ARCHITECTURE.md: renamed "Hourly NVIDIA NIM repair gate" → "Hourly contextual-orchestrator repair gate", updated its Mermaid diagram's NIM node to Gateway["contextual-orchestrator sidecar: orchestrator/free"], and corrected the "Control-plane data flow" sequence diagram and "Trust boundaries" bullet that both stated agents bind NVIDIA_NIM_API_KEY directly.
• CHANGELOG.md: ## [Unreleased] entry summarizing the correction.

Verification

• Targeted: python3 -m pytest tests/test_pr_review_autofix_nvidia_nim_contract.py tests/test_pr_review_fix_hourly_contract.py -q → 40 passed.
• Full suite: coverage run -m pytest tests && coverage report --show-missing && interrogate → 2126 passed, 1 skipped; coverage 100% (10543/10543 statements, 4228/4228 branches); interrogate 100%.
• Verified all contract-test-pinned substrings/anti-substrings in docs/doctoring/hourly-nvidia-nim-autofix.md survive the edit (ordinary and conflict repair, including ignored paths, `.git` and `.git/*`, `core.hooksPath=/dev/null`, explicit revalidated repository URL, the two forbidden legacy phrases, and the three required citations).
• git diff --check: clean.

No workflow, script, or test behavior changes — documentation only.

---

Generated by Claude Code

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[seonghobae]

exact-head-path-policy failed on this head (0c7803d3), job 99613849081: test_strix_quick_gate: 1 failure(s) — FAIL: opencode required workflow bootstrap must not depend on required-workflow event payload fields.

Not this PR's fault (this PR only touches ARCHITECTURE.md, CHANGELOG.md, and two docs/ files — nothing under .github/workflows/ or scripts/ci/). Root-caused it: scripts/ci/test_strix_quick_gate.sh:525's awk '/^ required-workflow-bootstrap:$/,/^[^ ]/' range has always had a boundary bug — /^[^ ]/ only matches a fully-unindented line, so it never stops at the next 2-space-indented job header and instead sweeps in every subsequent job in opencode-review.yml (coverage-source-tree, coverage-evidence, opencode-review-target) until end of file. That bug was latent until today's #1497 ("require substantive agent verdicts") added a genuinely new step-level if: github.event.action != 'closed' inside opencode-review-target — which the over-broad awk range now sweeps in and misattributes to required-workflow-bootstrap, even though that job itself has no if: at all and is fully compliant. Confirmed reproducing identically against a clean origin/main checkout (1cbb6aaf), independent of this PR's branch — this is base-branch-red, breaking exact-head-path-policy for every open PR in this repo, not just this one.

Fix ported: opened fix/strix-quick-gate-bootstrap-scope (targeting main directly, not this PR — the fix is unrelated to this PR's docs-only diff) with a corrected awk range that exits as soon as it reaches the next 2-space-indented job header, verified to (a) no longer match on the current opencode-review-target step and (b) still correctly catch a real violation if one were reintroduced. Full suite + this quick-gate script re-verified locally; will link the PR here once pushed. Not re-running this check on this head yet since the fix isn't merged — once it lands on main, the merge scheduler will update this PR's branch and this check will pass on the new head.

---

Generated by Claude Code

[seonghobae]

Fix pushed: #1528. Once that merges to main, the merge scheduler will update this PR's branch and exact-head-path-policy should pass on the new head.

---

Generated by Claude Code

[seonghobae]

Correction: #1528 (the fix I linked above) was closed by @seonghobae as a duplicate of the pre-existing, more thorough #1506, which already owns this exact fix (opened ~10 hours before I hit the same bug, and additionally catches a related awk | grep -q SIGPIPE-under-pipefail issue my narrower fix didn't). Once #1506 merges to main, this PR's exact-head-path-policy should pass on the next branch update — tracking #1506 instead.

---

Generated by Claude Code

[seonghobae]

Contextual-Orchestrator를 같이 손보든 어쩌든 해결하세요. Bypass merge 필요하면 가능 (chicken and eggs 상황이라면) + NVIDIA NIM 만 쓰는 건 허용하지 않아요. Contextual-Orchestrator를 쓰세요.

[seonghobae]

Contextual-Orchestrator와 관계한 것들을 같이 손보든 어쩌든 해결하세요. Bypass merge 필요하면 가능 (chicken and eggs 상황이라면) + NVIDIA NIM 만 쓰는 건 허용하지 않아요. Contextual-Orchestrator를 쓰세요. Timeout은 적어도 3시간으로 잡으세요. 120초 같은 건 당황스럽군요. Opencode와 Noema 는 Coderabbitai 및 Devin 수준으로 실제로 리뷰를 하게 하시오. Strix도 보안 리뷰를 꼼꼼하게 하도록 하시오. 특히 보안 리뷰는 전체 코드로 수행하는 것입니다. Contextual-Orchestrator는 실시간으로 빠르면서 능력이 좋은 모델에 요청을 보내어 시간을 당기시오.

[seonghobae]

Closing as semantically absorbed by protected main rather than restacking this 144-commit-behind documentation branch.

I re-read the live protected-main public/operational documentation before closing. ARCHITECTURE.md already names the Hourly contextual-orchestrator repair gate, renders the contextual-orchestrator sidecar: orchestrator/free gateway in the diagram, and states that the worker auto-discovers across the provider-secret registry rather than binding one provider directly. docs/automation/hourly-review-repair.md already says the write-capable worker uses OpenCode routed through the vendored contextual-orchestrator gateway and scopes the five provider secrets to sidecar provisioning. docs/doctoring/hourly-nvidia-nim-autofix.md already carries the dated 2026-08-31 correction, the orchestrator/free provider contract, and the corrected credential/rollback language. Those are exactly the four-document semantic requirements this PR was created to add; later main work has since evolved the same documents further.

Restacking this historical docs commit would therefore replay already-landed prose into newer architecture/operability material and create another full required-workflow fan-out without a remaining documentation gap. No unique current requirement is discarded; closing unmerged is the lower-risk queue-reduction action.