feat(agent): .hermes.md per-repository project config

[ch3ronsa]

Summary

Adds .hermes.md / HERMES.md per-repository project configuration — the #1 most requested UX pattern from the competitor analysis in #681. Every major AI coding agent has this (Claude Code's CLAUDE.md, Cursor's .cursorrules, Windsurf's .windsurfrules). Hermes now has it too.

When the agent starts a session, it discovers the nearest .hermes.md and injects its contents into the system prompt as project context — eliminating the need to re-explain project conventions every session.

How it works

my-project/
  .git/
  .hermes.md          ← discovered and loaded
  src/
    components/       ← agent started here, walks up to git root

Example .hermes.md:

---
model: claude-sonnet-4-20250514
tools:
  disabled: [tts]
---

# My Project

## Code Style
- Use TypeScript strict mode
- Prefer functional components with hooks

## Testing
- Run tests with `npm test`
- Minimum 80% coverage for new code

Changes

Modified: agent/prompt_builder.py — 3 new helper functions + integration into build_context_files_prompt()

Function	Purpose
_find_hermes_md(cwd)	Walk CWD → git root, return nearest .hermes.md or HERMES.md
_find_git_root(start)	Locate repo boundary (stops parent-dir walking)
_strip_yaml_frontmatter(content)	Remove --- delimited YAML block (structured config for future PR)

Modified: tests/agent/test_prompt_builder.py — 22 new tests

Design decisions

• Extends existing build_context_files_prompt() — same pattern as AGENTS.md and .cursorrules discovery. No new code paths.
• .hermes.md preferred over HERMES.md — dotfile convention (hidden by default), uppercase as fallback
• Nearest-first discovery — walks CWD → parent dirs → stops at git root. Subdirectory configs shadow parent configs.
• YAML frontmatter stripped — body-only injection into system prompt. Structured config (model/tools overrides) deferred to Phase 2.
• Injection scanner applied — same _scan_context_content() protection as all other context files
• 20K char truncation — same _truncate_content() limit as AGENTS.md

Scope (Phase 1)

This PR covers discovery & injection only. Future PRs can add:

• Frontmatter parsing for model/tool overrides
• .hermesignore for file/dir exclusion
• /project slash command to view/reload config
• Nested config merging (subdirectory extends parent)

Test plan

• 22 new tests (all passing), covering:
  • .hermes.md and HERMES.md discovery
  • Lowercase priority over uppercase
  • Parent directory walking up to git root
  • Stops at git root boundary (no leaking)
  • YAML frontmatter stripping
  • Injection blocking (prompt injection patterns)
  • Coexistence with AGENTS.md, .cursorrules, SOUL.md
  • Helper functions: _find_hermes_md, _find_git_root, _strip_yaml_frontmatter
• All 52 existing + new tests pass
• Manual: create .hermes.md in a repo, run hermes chat, verify rules appear in context

Closes #681 (Phase 1)

[teknium1]

Merged via PR #1712. Your .hermes.md discovery implementation was cherry-picked onto current main with authorship preserved — all 22 tests pass cleanly. Great work on the design (nearest-first discovery, git root boundary, YAML frontmatter stripping, injection scanning). Phase 2 (frontmatter parsing for model/tool overrides) can follow up separately.