fix: preserve all PRD sections during revision

[ekuris-redhat]

Summary

• Root cause: _strip_preamble checked text.find("\n#") before text.startswith("#"). When the LLM revision output started with ## 1. Executive Summary (no document title), the method found \n## 2. Problem Statement and stripped everything before it — including the entire Executive Summary.
• Fix: Swap the check order — startswith("#") first so content that already begins with a heading is never stripped.
• Also: Strengthened the shared regenerate.md prompt with explicit preservation rules for all content types (PRD, spec, epic, task).

Changes (3 files)

File	Change
src/forge/integrations/agents/agent.py	_strip_preamble: check startswith("#") before find("\n#")
src/forge/prompts/v1/regenerate.md	Added explicit "Preserve what was not objected to" instructions
tests/unit/prompts/test_prompt_templates.py	Regression test for preservation instructions

Test plan

• All 21 prompt template unit tests pass
• E2E on AISOS-742 (no new skill, _strip_preamble fix only):
  • Generated PRD → rejected with "change threshold to 500K"
  • Executive Summary: preserved
  • All 11 sections: intact
  • Only requested change applied: 50,000 → 500,000 (3 references, 0 old values remain)

🤖 Generated with Claude Code

[eshulman2]

fixed in #75

[ekuris-redhat]

Closing this PR — we were unable to consistently reproduce the Executive Summary deletion during PRD revision in local end-to-end testing.

While we did observe _strip_preamble stripping the Executive Summary once (when the LLM output started directly with ## 1. Executive Summary without a document title), the issue is non-deterministic and depends on the LLM's output formatting choices. Across ~15+ revision attempts, the bug only triggered once.

The root cause hypothesis (check order in _strip_preamble + weak regeneration prompt) remains plausible but needs more evidence before merging a fix. Will revisit if the issue is reported again with reproducible steps.