feat: add application learning queue for user feedback

[luochen211]

Problem

career-ops is designed to learn from user feedback after evaluations. Today, if the user says "this score is too high" or "you missed that I have experience in X", the agent is expected to update modes/_profile.md, config/profile.yml, or article-digest.md directly.

That is useful, but it has two risks:

• feedback may be applied too eagerly without user confirmation
• there is no audit trail of what the system learned and why

Proposed solution

Introduce a lightweight learning queue, for example data/learning-queue.md, where the agent records proposed profile updates as pending items:

• source interaction/date
• feedback summary
• proposed target file
• proposed wording/config change
• status: pending/applied/rejected

The agent can then ask for confirmation before applying each learning item to User Layer files.

Why this fits

It preserves the project rule that personalization belongs in the User Layer while making learning more reviewable and reversible.

Non-overlap audit

I checked open PRs before proposing this. This does not overlap with:

• Add recruiter-side application heuristics #839 recruiter-side application heuristics
• Add EU SWE application calibration mode #840 EU SWE application calibration
• feat(cv): add fact validator for generated resumes #822 CV fact validator

This is about user-feedback workflow and auditability, not adding generic scoring heuristics.

Suggested implementation scope

• Define data/learning-queue.md format
• Add mode guidance in modes/_shared.md or a small docs page
• Optionally add learning-queue.mjs for list/apply/reject operations
• Keep all actual personalization writes behind explicit user confirmation

Test plan

• Markdown format documented
• Optional script self-test if implemented
• Existing data contract continues to treat data/* as User Layer

[santifer]

Closing with its PR under the core-acceptance criterion explained on #890 (core takes what the candidate uses; project-artifact tooling lives outside the core — companion-repo door open).