df: wire --security security-review mode into the review commands (spec 010)

[bborbe]

Wires the --security security-review mode into /coding:pr-review, /coding:code-review, /coding:local-review (spec 010).

• Prompt 042: pr-review.md — canonical full security pipeline (recon → /tmp/security-model.json, six trait groups + authz over-selection + deterministic invariant selection, adjudicator inputs, verifier gate, derived blocking, PR-mode diff anchoring, Security Findings + Security Model report sections, SECURITY_MODEL_FILE citation validation)
• Prompt 043: code-review.md — audit-mode mirror (whole-repo scope, no diff anchoring)
• Prompt 044: local-review.md — local-mode mirror (HEAD~1 diff anchoring)
• Prompt 045: scenarios 007/008/009 finalized (TBD fixture blocks → inline-built generic fixtures + /coding:local-review --security walks) + new 010 security-zero-findings
• Prompt 046: README commands/Quick Start/Acceptance-Scenarios rows, llms.txt alignment, CHANGELOG ## Unreleased feat entry

Frozen contract untouched: docs/selector-mode-guide.md (dormant Security Extension), scripts/validate-citations.sh, agents/security-verifier.md, agents/go-security-specialist.md, docs/security/* (AC6 enforced via checksum baseline).

[ben-s-pull-request-reviewer]

All checks pass. Let me compile the final findings and produce the verdict JSON.

Plan concern verification summary:

Concern	Verified?	Evidence
SECURITY_REVIEW=1 gates dormant extension	✅	argument-hint updated in all 3 commands; SECURITY_REVIEW=1 set on flag detection
blocking formula derived (not stored as severity)	✅	All 3 commands: blocking = confidence==confirmed ∧ exploitability==high ∧ impact≥medium; explicit "never stored as severity"
/tmp/security-model.json session-local	✅	All 3 commands: "never to any path inside the reviewed repo — if accidentally written in-tree, delete and rewrite"
SECURITY_MODEL_FILE citation validation fail-closed	✅	All 3 commands: SECURITY_MODEL_FILE=/tmp/security-model.json passed to validator; absent/unreadable model → WARN: dropped fail-closed
HARD INVARIANT preserved	✅	All 3 commands: "The HARD INVARIANT holds: applicable set is a subset of Step 4b-i candidate set; trait groups never add a rule the glob did not produce"
Toolchain pass fail-closed	✅	All 3 commands: "A scan failure, a database-fetch timeout, or a flagged vulnerability surfaces as a Must-Fix toolchain finding — never a silent skip"
--security × short mode: pipeline runs even when short skips Step 4	✅	pr-review: line 85; local-review: line 25; explicit "never silently ignored"
--security × --refresh-baseline: maintenance exit not silent ignore	✅	code-review line 33: "exits WITHOUT a report — on such an invocation no security pipeline runs, because there is no review being performed; this is not a silent ignore of the flag"
audit scope (code-review) is whole-repo, no baseline suppression	✅	code-review line 195: "baseline-independent whole-repo inventory"
4 scenarios: all have ≥8 checkboxes, 0 TBD, each has --security invocation	✅	Grep: TBD count = 0 for 007/008/009; checkbox counts = 18/14/17/13 ≥8 each

AC6 (negative guard): git diff origin/master -- docs/selector-mode-guide.md scripts/validate-citations.sh agents/security-verifier.md agents/go-security-specialist.md docs/security/ → empty ✅

No code findings — this is a procedure/spec wiring PR for a coding plugin; changed files are commands/docs/scenarios/prompts/specs with no executable code in scope.

{
  "verdict": "approve",
  "summary": "PR 122 wires the `--security` flag into all three review commands, activating the dormant security extension. All 10 plan concerns are correctly addressed: SECURITY_REVIEW=1 gates the pipeline, blocking is derived from the verifier verdict formula (not stored as severity), the security model is session-local at /tmp/security-model.json, citation validation is fail-closed via SECURITY_MODEL_FILE, the HARD INVARIANT is preserved (applicable ⊆ Step 4b-i candidates), the toolchain pass is fail-closed, short mode cannot suppress the security pipeline, --refresh-baseline maintenance exit is documented as non-silent, audit scope is whole-repo and baseline-independent, and all 4 scenarios have ≥8 checkboxes, 0 TBD blocks, and the --security invocation. AC6 negative guard confirmed: shipped artifacts (selector-mode-guide, validate-citations, security-verifier, go-security-specialist, docs/security/) are untouched.",
  "comments": [],
  "concerns_addressed": [
    "security: SECURITY_REVIEW=1 gates dormant security extension in all three commands — verified by argument-hint and flag detection text",
    "security: blocking formula (confidence==confirmed ∧ exploitability==high ∧ impact≥medium) derived at render time, never stored as severity — verified in all three command files",
    "security: /tmp/security-model.json is session-local, never written in-tree; accidental in-tree writes are deleted and rewritten — verified in all three command files",
    "security: SECURITY_MODEL_FILE=/tmp/security-model.json passed to validate-citations.sh; absent/missing model → invariant findings drop fail-closed with WARN — verified in all three command files",
    "correctness: HARD INVARIANT preserved — applicable set is subset of Step 4b-i candidate set; trait groups never add a rule the glob did not produce — verified in all three command files",
    "security: toolchain pass is fail-closed — scan failure, DB timeout, or flagged vuln surfaces as Must-Fix, never silent skip — verified in all three command files",
    "correctness: --security × short mode — security pipeline runs even when short mode skips Step 4; flag is never silently ignored — verified in pr-review.md:85 and local-review.md:25",
    "correctness: --security × --refresh-baseline — command exits after writing baseline without running pipeline; documented as maintenance-mode exit, not silent ignore — verified in code-review.md:33",
    "security: audit scope (code-review --security) is whole-repo; Security Findings are baseline-independent and not suppressed by severity filter — verified in code-review.md:195",
    "tests: 4 acceptance scenarios (007/008/009/010) — all have ≥8 unchecked checkboxes (18/14/17/13), zero TBD blocks, each names exact --security invocation — verified via grep and file reads"
  ]
}