Open alpha: collect precision and noise feedback from real PR reviews

[direct41]

Problem

The deterministic public corpus is useful as a regression contract, but curated retrospective evidence does not replace feedback from engineers using the output during real review work.

Goal

Run a limited open-alpha feedback cycle focused on whether the top three findings change reviewer or test behavior.

Scope

• Collect at least five independently submitted or reviewer-observed public PR runs.
• Prefer backend-heavy FastAPI, Django/DRF, and Express/Node changes.
• Record the command, merge decision, top three findings, useful output, noise, missed risk, setup friction, and workflow preference.
• Separate product defects from unsupported-stack requests and distribution requests.
• Turn repeated precision/noise patterns into focused follow-up issues or code changes.
• Publish only privacy-reviewed aggregate conclusions.

Guardrails

• Do not include private repository names, customer context, secrets, or non-public vulnerability details.
• Do not count internally curated benchmark labeling as external alpha feedback.
• Do not prioritize hosted workflow or broad stack expansion until repeated requests justify it.

Done when

• At least five real PR review feedback records are collected.
• Top-three precision/noise and reviewer/test impact are summarized.
• Setup blockers and requested workflows are categorized.
• Repeated product defects have follow-up issues or fixes.
• A continue/change/pause decision for the alpha workflow is documented.

[direct41]

Autonomous readiness work completed in #81.

• Alpha feedback now captures reviewer role, merge decision, top three findings/actions, review impact, setup friction, workflow preference, repeat intent, and privacy acknowledgement.
• The public validation docs and result template match those fields.
• Clean wheel install and bundled sample smoke passed.
• The 49-case public corpus was rerun; three open PR heads had changed and were re-reviewed.
• Corpus labels are now pinned to reviewed head SHAs, and future head drift is reported explicitly instead of looking like a product regression.
• Full quality, package, public-artifact, dependency, CodeQL, Dependabot, and secret-scanning checks are clean.

Remaining blocker for #72 is external evidence: at least five independently submitted or reviewer-observed real PR feedback records.