ML-196: add application-task pi prompt template

Introduces /application-task for working on Backlog.md tasks end-to-end:
- Pre-flight: read docs, load task plan, load relevant skills
- Execution: present plan, short loops, log progress, handle scope
- Done vs Complete: finalize without archiving, per Backlog.md workflow

.pi/prompts/application-task.md

@@ -0,0 +1,68 @@
+---
+description: Work on an existing backlog task end-to-end
+argument-hint: "<TASK-ID>"
+---
+
+We're going to work on task $1 end-to-end. Follow the Backlog.md Task Execution workflow
+step by step, ensuring every step is completed before moving to the next.
+
+## Pre-flight
+
+Before touching any code:
+
+1. **Read the project documentation.** Open and read `docs/architecture.md` and
+   `docs/project-conventions.md`.
+
+2. **Load the task plan.** Use `backlog_task_view` to load task $1 and read its
+   implementation plan, acceptance criteria, and notes. Do not proceed until you
+   understand the plan.
+
+3. **Load relevant skills based on the task's domain upfront.**. When undecided
+   if a skill is relevant or not, opt for loading it anyway.
+
+## Execution
+
+Follow the Backlog.md Task Execution workflow:
+
+1. **Mark the task In Progress** and assign yourself via `backlog_task_edit`.
+2. **Present the plan to the user** — summarize the approach and ask for
+   confirmation before writing any code. Wait for explicit approval.
+3. **Work in short loops:** implement a step, run the relevant tests, and
+   immediately check off acceptance criteria with `backlog_task_edit`
+   (`acceptanceCriteriaCheck` field) when they are met. If you catch yourself
+   running in circles, stop and let the user know.
+4. **Log progress** with `backlog_task_edit` (`notesAppend` field) to document
+   decisions, blockers, or learnings.
+5. **If the plan needs to change**, update it first via `backlog_task_edit`
+   (`planSet` or `planAppend`), get confirmation, then continue.
+6. **If new work appears outside the original acceptance criteria**, STOP and
+   ask the user before expanding scope. Never silently add AC or create
+   follow-up tasks.
+
+## Done vs Complete
+
+When implementation is finished, do NOT call `backlog_task_complete` or
+`backlog_task_archive`. Those move the task to the archive — they are for
+periodic batch cleanup, not for marking work finished.
+
+Instead, follow the Task Finalization workflow:
+
+1. **Verify all acceptance criteria** are checked (use `backlog_task_view` to
+   review).
+2. **Verify all Definition of Done items** are checked.
+3. **Confirm tests pass** — run the full test suite or the relevant test files
+   and verify no new warnings or regressions.
+4. **Write the Final Summary** via `backlog_task_edit` (`finalSummary` field).
+   Include what changed, why, tests run, and any risks or follow-ups.
+5. **Update the plan** if the executed approach deviated from what was recorded.
+6. **Set status to "Done"** via `backlog_task_edit`.
+
+Only then is the task truly done.
+
+## After Finalization
+
+- **Never autonomously create or start new tasks.**
+- If follow-up work is needed, present the idea to the user and ask.
+- If this was a subtask and the user instructed you to work on a parent "and all
+  subtasks," proceed to the next subtask. Otherwise, ask the user whether to
+  continue.