dashboard: say how fresh it is, and hand over its own refresh

[Jammy2211]

Merge this before the PyAutoMind twin (PyAutoLabs/PyAutoMind#…, same branch name). dashboard_refresh.yml renders the Mind page with PyAutoBrain's default branch, so if the Mind side lands first its PR check reads the new banner as content drift and the self-heal on main commits it straight back out.

Why

The Mind dashboard's failure mode is asymmetric. dashboard_refresh.yml self-heals a stale render on every push to main, but nothing self-heals a stale prompt: a task that ships without its prompt advancing to complete/ keeps rendering as pickable backlog, and a human finds out by starting work that is already done. That happened often enough to be the ask behind this change.

What changed

A freshness banner on both twins, above the counts:

• the generation date;
• one paragraph on why a fresh render is not a fresh backlog;
• and — in the same one-tap 📋 idiom as every task row — the whole reconcile-then-regenerate chore as a single copyable instruction (pull, run the lifecycle checks, decide which prompts are already done and why, write each record and retire its prompt, regenerate, push).

The markdown page gets a <details> row; the HTML twin gets a real copy button and its own .fresh rule, since no other organ board carries one.

A widened drift detector. _FIX_PR_RE caught only the Fix: … PR #n idiom. The commoner signal is the prompt's own Status: header — the shipping session writes shipped / superseded / absorbed into it and leaves the file in draft/. _done_status() reads that off the parsed header, and only when the status opens on a done-word, so a tracker reporting phases 1-3 SHIPPED; phase 4 open stays live work. Run against PyAutoMind it immediately surfaced two spent prompts a manual sweep had missed.

Validation

455 passed — the full Brain suite. New contract tests cover: shipped/superseded/absorbed statuses flag; partly-shipped ones do not; both twins carry the banner; and the HTML blurb renders its code spans while the copy payload keeps its markdown verbatim.

That last one needed a scoping fix to a pre-existing test. test_the_html_blurb_renders_its_code_spans asserted over the whole page, and a copy payload is a clipboard literal — the message a human pastes into a Claude chat — so it legitimately carries markdown the page must not render. A shared _prose() helper strips data-cmd attributes, which is what that assertion was always about.

---

Generated by Claude Code