feat: real agentic Propose via GitHub Copilot CLI, remove mock

[ineedjet]

Summary

Splits Propose into a deciding half and a mechanical half, mirroring the split #11 already draws between Apply/Reject (deterministic) and Distill/Rework (agent work):

• New propose-context mode: gathers regulations text, the candidate issue's title/body, and the field catalog into one self-contained prompt. Output only — agnostic to whichever inference mechanism reads it.
• propose now requires --model-response (JSON: {"title", "body", "fields": {...}}) instead of deciding content itself. mock_value_for is deleted entirely — mock mode is no longer needed, per instruction.
• proposal-shared.yml's propose job now runs propose-context → installs Copilot CLI → calls actions/ai-inference with that prompt → feeds the response into propose.
• validate hard-errors on schedule/workflow_dispatch runs missing model_credentials, same as it already does for capacity_query/queue_query.

A real detour: GitHub Models is retired

Went looking for "GitHub's own built-in agent tooling" and found GitHub Models (the inference API this would have used) was fully retired July 30, 2026 — playground, catalog, inference API, BYOK, gone for everyone. actions/ai-inference pivoted to wrap GitHub Copilot CLI instead, which is what this PR actually uses. Requires a PAT with an active Copilot seat, not GITHUB_TOKEN.

A real bug found along the way

write_output wrote plain key=value\n lines to $GITHUB_OUTPUT — corrupts the file for any multiline value. prompt is the first output that's always multiline. Fixed by switching to GitHub's documented delimiter form unconditionally.

Dogfood tuning (this repo's own regulations_path: TRANSLATE.md)

• Test Project retitled to Translate.
• Replaced the five generic placeholder fields (Summary/Points/Due/Priority/Tags) with a single real Proposal Size/Size pair (SINGLE_SELECT, options matching this repo's own □ XS/S/M/L/XL label names — reused for a familiar scale, not read from those labels; the regulation estimates size from text density, not effort).
• proposal.yml wired to a new COPILOT_PAT secret — not yet created, needs a PAT with an active Copilot seat before Propose can run for real here.

Verification

Live against real repository data, not assumed:

• propose-context produces a correct, complete prompt — regulations text, a real non-English test issue's title/body, and the live field catalog.
• propose --model-response correctly creates a real proposal from a hand-crafted JSON response — title, body, and the Size field all landed correctly. Confirms the mechanical half works end-to-end.
• The real model call (actions/ai-inference → Copilot CLI) still needs COPILOT_PAT to exercise for real — not yet tested live.

Test plan

• pre-commit run --all-files passes.
• propose-context verified live (real regulations file, real non-English issue, real field catalog).
• propose --model-response verified live end-to-end (hand-crafted JSON → real created proposal, all fields correct).
• Full real model call through actions/ai-inference/Copilot CLI — blocked on COPILOT_PAT secret creation.

Related: #11, #9

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: dac6c6eafe

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Validate model fields before creating the issue

When inference returns valid JSON whose fields value is an array or string, the proposal has already been created before this .items() call raises AttributeError; it is therefore left without field values, a control comment, or its transition, and a retry can create a duplicate proposal. Validate that the response is an object and fully normalize its fields before performing createIssue.

AGENTS.md reference: AGENTS.md:L138-L138

Useful? React with 👍 / 👎.