feat(prompts): Kiro-style agent identity + explicit edit contract

[JessicaMulein]

Summary

Rewrites the agent-family system prompts so models follow the editing contract up front instead of learning it only after a failed edit. The previous agent prompt was a thin directive list that (a) encouraged long exploration loops, (b) never stated the #1 cause of failed turns (edit ordering), and (c) had no identity/voice.

Changes

• prompts/agent.yml main_system — expert-engineer identity; investigate-before-claiming; scope discipline; failure-loop recognition (stop after two failures, diagnose); and an explicit editing contract: ContextManager -> ReadRange -> EditText, one file per EditText call, @000/000@ markers for empty files. Drops the loop-encouraging 'no task takes too long' line.
• Bugfix: the agent prompt never referenced {final_reminders}, so overeager_prompt and the MCP tool_prompt were silently dropped for the agent coder. Now referenced exactly once.
• prompts/subagent.yml — now inherits the agent identity/contract instead of re-overriding main_system with stale directives; keeps only the sub-agent-specific verbose Yield summary.
• prompts/ask.yml / prompts/architect.yml — same direct voice and ground-answers-in-code discipline; architect plans now name verification steps and edge cases.

Test plan

• New tests/basic/test_a-ent_prompt_contract.py (deterministic, no LLM): prompts render via str.format with no stray braces and no unknown keys; {final_reminders} appears exactly once across main_system+system_reminder; edit contract is stated (ReadRange before EditText, one file, empty-file markers); sub-agent inherits the agent main_system; legacy 'no task takes too long' directive is gone.
• pytest tests/basic/test_prompts.py — 27 pass (inheritance chains unchanged).
• pytest tests/coders/test_copypaste_coder.py tests/subagents/ — pass.
• Behaviorally validated end-to-end against a real local model (qwen3-coder:30b) via the consuming app's harness: scoped edit task completed, zero edit failures, ReadRange preceded the edit.

Notes for reviewers (scope & how to split)

These changes fall into two buckets, and I'm happy to split if you'd prefer to land them separately:

• Objective fixes (recommend landing regardless):

  • {final_reminders} was never referenced in the agent prompt, so overeager_prompt and the MCP tool_prompt were silently dropped for the agent coder. Now referenced exactly once.
  • subagent.yml re-overrode main_system with a stale copy, so sub-agents diverged from the main agent every time it improved. Now inherits.
  • The editing-contract block restates rules that come straight from cecli's own edit_text.py / read_range.py error messages (ReadRange-before-EditText, one file per call, @000/000@), so it should reduce failed turns broadly — most on smaller/local models.

• Opinionated / negotiable:

  • The identity/voice rewrite is a matter of taste.
  • main_system got longer (~2x); for tiny-context local models that's a real token cost. Happy to trim.
  • I removed the "no task takes too long / be persistent" line because it encouraged exploration loops in our headless harness. A pure-CLI user may prefer to keep it — call it and I'll restore.

If you want only the objective bucket, I can drop the persona/voice + length changes into a follow-up and keep this PR to the two bugfixes + the contract block.

Honest scope of validation: the new test proves the prompt states the right things; behavioral improvement was validated end-to-end on one model (qwen3-coder:30b) on one scoped edit task (completed, zero edit failures, ReadRange preceded the edit). It is not a multi-model benchmark.