fix(pr-af): enforce canonical severity vocabulary + cross-project HITL callback

[AbirAbbas]

Summary

A real review with 3 findings on a PR (Agent-Field/github-buddy#73) found genuine bugs but posted nothing to GitHub. Root cause turned out to be two stacked issues; this PR fixes both.

Root cause

1. Severity vocabulary mismatch (why nothing posted). PR-AF gates posting behind a HITL approval request (hax-sdk pr-af-review-v1). One finding came back with severity: "high". The hax-sdk template enforces a zod enum {critical, important, suggestion, nitpick}, so the approval-request create returned 422 (Invalid enum value … received 'high'). request_review_approval treats any create failure as REJECT ("never post an unreviewed review") → the whole review was dropped, and the run still "succeeded".
   • It slipped through every guard because PR-AF typed severity as a bare str (the allowed set was only a comment). So the model's "high" was a valid string to PR-AF's pydantic schema and to the SDK harness's schema-validation/retry loop — there was no validation failure to retry on. The enum lived only at the hax-sdk boundary, which has no retry/feedback loop back to the model.
2. Cross-project callback (why the morning's answered review never resumed). approval_webhook_url() built the callback from AGENTFIELD_SERVER (the internal control-plane.railway.internal address the agent uses to reach the CP). hax-sdk runs in a separate Railway project, so the webhook fetch failed and the response never got back.

Fixes

• schemas/severity.py — single source of truth. Severity = Annotated[Literal[...], BeforeValidator(normalize_severity)]: the JSON schema handed to the model now advertises the enum, and stray labels are coerced (high→important, medium→suggestion, low→nitpick) instead of failing — deterministic, with no retry storms or dropped findings.
• Applied to ReviewFinding, ScoredFinding, _CompoundFinding, CrossRefInteraction; revised_severity normalized at consumption (model_copy bypasses validators); normalized once more in _finding_payload as a last line of defense before the hax create.
• Reviewer prompts now state severity MUST be exactly one of the four values.
• Surfaced the silent failures: the HITL "not posting" log now includes the underlying create error, and a total schema-parse failure in review_dimension no longer masquerades as "0 findings".
• Cross-project callback: approval_webhook_url() now prefers AGENTFIELD_PUBLIC_URL (public, cross-project reachable), falling back to AGENTFIELD_SERVER for single-project/local setups — keeping agent→CP traffic internal.

Test Plan

• ruff check src/ scripts/ clean (the CI lint gate)
• pytest — 56 passed (py3.12), incl. new test_severity.py (normalize + model coercion + payload boundary) and approval_webhook_url precedence tests
• Verified the generated JSON schema for severity now emits enum: [critical, important, suggestion, nitpick]

Companion ops change (not in this PR)

Set AGENTFIELD_PUBLIC_URL=https://<control-plane-public-domain> on the pr-af Railway service so the callback fix takes effect in production. (AGENTFIELD_SERVER stays internal.)

🤖 Generated with Claude Code