agency: drop per-thread refine-context file; query the DB instead

[MagMueller]

Why

The Edit (refine) flow used to write the original card's title + description + prompt to `/var/lib/bux/agency-refine-context/.txt` on tap, then `run_task` would read + delete that file on the user's first reply to seed the worker agent with the original context.

Two redundant surfaces — the data was always already in the suggestion row (`title`, `description`, `prompt`). The file was a duplicate-state cache that needed its own ownership / permissions handling on `/var/lib/bux`.

What

`agency_db.pop_refine_context_for_thread(db, thread_id) -> str | None` — atomic SELECT + UPDATE. Returns the assembled plain-text context AND flips a new `refine_context_injected` column 0 → 1 in one call. The context is injected exactly once, on the user's first reply in the worker topic. Subsequent replies in the same thread are normal lane runs.

Schema migration via `ALTER TABLE` with duplicate-column swallow, mirroring how `spawn_topic` was added in #107. Pre-existing rows get `refine_context_injected = 0` by default.

`run_task` swaps the file pop for the DB call. The kind=refine handler stops calling the file-writer. The file-based helpers (`_agency_write_refine_context`, `_agency_pop_refine_context`, `_agency_build_refine_context_plain`, `_AGENCY_REFINE_CTX_DIR`) are removed — `agency_db` owns the plain-text assembly now. The visible HTML context message in the new topic still renders via `_agency_build_refine_context`, just with the new `📎 Context` emoji + title (from #111).

Net

One fewer state surface, one fewer ownership/permissions race, the same UX for the user (Edit → context shown → "what would you change?" → user replies → agent re-drafts with original in scope).

Test plan

• Schema migration runs cleanly on the existing DB; `refine_context_injected` column appears
• `pop_refine_context_for_thread` on a non-worker thread → None
• Tap Edit on a card → context message appears in spawned topic, no file created at `/var/lib/bux/agency-refine-context/`
• User replies in the spawned topic → agent fires with original context prepended
• User replies again → agent fires normally (no re-injection of context)

---

Summary by cubic

Replaced the per-thread refine-context file with a DB lookup that injects the original card context once on the user’s first reply. This removes duplicate state and permission hassles while keeping the same UX.

• Refactors
  • Added refine_context_injected column (default 0) with duplicate-column-safe migration.
  • Implemented agency_db.pop_refine_context_for_thread(db, thread_id) to assemble context and flip the flag atomically.
  • Updated run_task to use the DB call; removed _agency_write_refine_context, _agency_pop_refine_context, _agency_build_refine_context_plain, and _AGENCY_REFINE_CTX_DIR.
  • Kept the visible context message via _agency_build_refine_context with the new "📎 Context" label.

^{Written for commit f6a1719. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 2 files