fix(scan): use the registered scan ID in agent prompts

[dipeshbabu]

Fixes #290
Fixes #292

Summary

• Build the agent prompt only after the workbench returns the authoritative scan ID.
• Embed that ID directly in standard-scan artifact guidance and deep-scan MCP arguments.
• Keep the corrected prompt when false-positive feedback is appended.
• Cover standard and deep modes with and without feedback.

Root cause

The prompt was created before scan registration and referred to $CODEX_SECURITY_SCAN_ID. That placeholder is not expanded inside structured MCP arguments or agent-authored artifacts, so deep scans could send an invalid literal ID and standard scans could write a different UUID that finalization rejected.

Impact

Deep scans now call start_codex_security_deep_scan with the registered UUID, and all scan modes instruct the agent to use that same UUID in canonical artifacts.

Validation

• pnpm dlx bun test --timeout 30000 ./tests-ts/api.test.ts --test-name-pattern "uses the registered scan ID in standard and deep prompts|encodes valid Unicode Git refs as data before sending the scan prompt|provides only reviewed false positives" — 3 passed
• pnpm exec tsc --noEmit
• pnpm exec prettier --check src/api.ts tests-ts/api.test.ts
• git diff --check

The complete API test file reached 80 passing tests; its remaining nine tests are blocked in this local Windows environment by the existing credential-home ancestor ACL check.

[mldangelo-oai]

Thanks for tracking this down and contributing the fix, @dipeshbabu, and thanks to @charle-z for the earlier investigation. I carried your original commits into #354, merged current main, preserved both contributors' credit, and added a small regression for scan-skill preflight. Our maintainer tooling cannot push directly to personal forks, so the follow-up lets the normal CI run without losing your work or authorship.

[mldangelo-oai]

Thank you again for tracking this down and contributing the fix, @dipeshbabu! Your original commits and authorship were preserved in #354, which has now merged and resolved both linked issues. I am closing this PR because your fix has already landed through that contributor-preserving follow-up. We really appreciate your help improving Codex Security!