feat: configurable working-notes & cluster LLM prompts

[kastriotkastrati]

What

Makes the two LLM summary prompts configurable via env, with verbatim fallback to the current prompts.

• Working notes (per-thread GitHub sticky comment) → NOTES_PROMPT
• Cluster summary (cross-thread rollup) → CLUSTER_PROMPT

How

• System/user split: the configured text is sent as the model's system prompt; the thread data (title/state/description/discussion) stays the user message, assembled in code. Truncation bounds, bot/own-comment exclusion, and output sanitization remain in code and are not overridable.
• Fallback: unset/blank → built-in DEFAULT_NOTES_PROMPT / DEFAULT_CLUSTER_PROMPT, which are the original prompts verbatim. An unset secret behaves exactly as before.
• Idempotency: both prompts fold into the notes/cluster content hash, so editing a prompt re-renders existing notes and busts the AI Gateway cache. promptVar maps blank/whitespace env → undefined so an empty prompt is unrepresentable (schema .min(1).default(...)).

Updating in production

Cloudflare secrets, no redeploy (takes effect next event):

cd apps/worker
wrangler secret put NOTES_PROMPT < notes-prompt.txt
wrangler secret put CLUSTER_PROMPT < cluster-prompt.txt

See docs/configuring-prompts.md.

Verification

typecheck 13/13 · test 13/13 · lint clean (incl. no-raw-loops). Merged latest main (raw-loop refactor) into the branch.