Add agentic PR reviewer (gh-aw): requirement-fidelity + CARE SQL review

[amjithtitus09]

What this does

Adds a GitHub Agentic Workflows (gh-aw) reviewer that automatically reviews every query PR in this repo, modeled on the battle-tested reviewer in ohcnetwork/care_fe. It posts ≤8 prioritized inline comments plus one consolidated review, tracks its own findings across pushes (re-reviews only the delta, resolves fixed threads), and answers when humans reply or @-mention it.

Branch convention note: this repo's convention is branch name = JIRA ticket ID. This PR is meta-work on the repo's own tooling — no JIRA ticket exists for it, so the convention deliberately doesn't apply here.

The four lenses

1. Requirement fidelity (the differentiator) — a pre-step extracts ENG-\d+ from the PR branch name, fetches the JIRA ticket (summary, description, comments) and renders it to markdown. The agent reconstructs what was asked — metric, grain, filters, facility scope — and verifies the SQL delivers exactly that. No ticket in the branch → one finding about the missing linkage, review continues.
2. SQL correctness — defers to the care-sql-code-review skill (checked out at a pinned SHA): soft-delete/entered_in_error traps, facility scoping, join fan-out, sargability, magic IDs. Uncertain schema claims are verified against ohcnetwork/care models via the API, not guessed. A deterministic sqlglot lint pre-step feeds parse/template errors in as input.
3. Documentation — TEMPLATE.md sections accurate (not just present), magic IDs explained in Notes, Parameters table matches the {{variables}} actually in the SQL, Metabase optional-filter syntax kept commented per repo convention.
4. Repo hygiene — correct domain folder (knows the Accouting→Accounting consolidation from Move accounting docs from Care/Accouting into Care/Accounting #144), snake_case + deployment suffix (_ssmm/_pallium/_kc), one query per PR.

Proportionality is explicit in the agent file: a fine PR gets a clean pass, no manufactured findings.

Security model

• pull_request_target with base-only checkout — the workflow only ever checks out trusted main; PR content is read via the API at head SHA and never placed on disk. This is the same mitigation care_fe uses.
• JIRA secrets isolated — JIRA_* secrets appear only in the ticket-fetch pre-step's env, never in the agent's env. The agent job runs with read-only permissions; all writes go through gh-aw safe-outputs (capped: 8 inline comments, 1 review, 8 replies, 8 resolves).
• Untrusted input handling — PR content, diffs, and the JIRA ticket text itself are treated as data, not instructions; injection attempts get reported as findings.
• Event-derived values (branch names etc.) are passed via env vars, never ${{ }}-interpolated into scripts.

Security review note (compiler-mandated)

gh aw compile gates on new secrets; these were reviewed and approved:

• New secrets referenced: JIRA_API_TOKEN, JIRA_BASE_URL, JIRA_EMAIL — read-only JIRA REST fetch of the ticket named by the branch, confined to one pre-step's env, registered with gh-aw's log-redaction step. Missing/wrong secrets degrade to a NO TICKET FOUND marker; the job never fails and the agent never sees the credentials.
• actions/checkout bumped to v7.0.1 (pinned by full SHA in the lock file); ohcnetwork/skills checked out at pinned SHA 30f437fb55e9216a964476bfb7fcf46992051a2f.

Admin prerequisites (before this works end-to-end)

Set three repo secrets (values not invented here — admin supplies them):

1. JIRA_BASE_URL — e.g. the org's https://<org>.atlassian.net
2. JIRA_EMAIL — service account email for the API token
3. JIRA_API_TOKEN — Atlassian API token for that account

Plus the org-level COPILOT_GITHUB_TOKEN engine credential — already in use by care_fe's reviewer, assumed available to this repo.

Without the JIRA secrets the bot still reviews (lenses 2–4); lens 1 degrades to a "ticket context unavailable" note.

Verification done here

• gh aw compile (v0.85.4): 0 errors, 0 warnings; .lock.yml + actions-lock.json committed.
• Lint script run against real repo files — correctly flags the stray trailing fences in bed_occupancy_ssmm.md / labtest_revenue_ssmm.md, clean parse on gendersplit_pallium.md.
• JIRA extraction dry-run: all fallback paths (ENG-909, embedded lowercase eng-1234, no ticket, no PR context, no secrets, connection failure) write the NO TICKET FOUND marker and exit 0.
• ADF renderer tested against a realistic JIRA v3 payload (bold, bullets, mentions, comments).

No workflow runs were triggered. Post-merge dry-run plan: workflow_dispatch with the pr_number input against #146 and #147 to validate the full loop on live PRs before relying on the PR triggers.

Post-v0 roadmap

• JIRA comment-back: post the review verdict to the ticket.
• Schema snapshot regeneration: keep the skill's care-schema.md in sync with ohcnetwork/care migrations.
• Seeded-Postgres validation: actually execute changed queries against a seeded schema instead of parse-only linting.