Blind screening: redact identity fields (name/location/company) from the evaluator prompt

[AdvancedUno]

Problem

Fairness is currently only a prompt instruction ("NEVER let name, gender, race… affect any score"). Instructed LLMs still exhibit measurable demographic score deltas, and we feed the model exactly the signals that carry those cues: the candidate's name (ResumeParsed.name, GitHubSignal.name), location (GitHubSignal.location), and company. _github_summary puts name= straight into the dossier. Telling the model to ignore what we hand it is not a control.

Proposal

Implement blind screening: strip personal-identity fields from a copy of the profile before building the LLM dossier. The original profile is untouched and still drives the report's Candidate: header — only the prompt is redacted.

Redact (demographic cues, not needed to score skill):

• ResumeParsed.name → blank
• GitHubSignal.name / .location / .company / .twitter_username → None

Preserve (professional signal, not demographic):

• GitHubSignal.username, .blog, work-history company names, discovered.github, discovered.email.

Implementation notes

• New _redact_for_prompt(profile) -> CandidateProfile in evaluator.py using model_copy(update=...).
• Call it as the first line of _user_prompt.
• Note: resume.text still contains the name/email inline; fully scrubbing free text is out of scope here (removes signal, high false-positive). This issue targets the structured fields, which is where _github_summary/_resume_structured_summary surface identity most cleanly. A follow-up can consider free-text redaction.

Acceptance criteria

• Candidate name, GitHub location, and GitHub company do not appear in the structured dossier sections of the prompt.
• GitHub username still appears (professional identifier).
• The rendered report still shows the candidate's real name in the header.
• Test in tests/test_evaluation.py asserting the redaction.

Files

• src/hiregauge/evaluator.py
• tests/test_evaluation.py

Blocks: counterfactual fairness tests (separate issue). Related: fairness prompt in evaluator._FAIRNESS.

[AdvancedUno]

Regression guard for this redaction is tracked in #42 (counterfactual name-swap tests). Implement this first — #42 depends on it.