fix(integration): respect comma-separated applyTo globs in Claude/Cursor/Windsurf

[danielmeppiel]

Fixes #1366

WHY

applyTo frontmatter is documented as "Glob (or comma-separated globs)" in
docs/src/content/docs/producer/author-primitives/instructions-and-agents.md,
but the Claude, Cursor, and Windsurf converters emit comma-list values as a
single literal glob string, and the context optimizer feeds the unsplit value
to glob/fnmatch (producing misleading Pattern '...' matches no files warnings).

Copilot was already correct because its converter preserves applyTo verbatim
and the consuming tool splits — that path is intentionally unchanged.

WHAT

• New helper parse_apply_to(value) -> list[str] in src/apm_cli/utils/patterns.py
  owns the canonical comma-split semantics (whitespace trim, empty / trailing /
  leading / lone-comma tolerance).
• _convert_to_cursor_rules, _convert_to_windsurf_rules, and
  _convert_to_claude_rules now call the helper and branch:
  • single glob → existing scalar form (backwards compatible)
  • multi glob → YAML list under the target's native key
• context_optimizer._file_matches_pattern splits only top-level commas
  (so brace alternation like **/*.{css,scss} is unaffected) and treats any
  segment match as a hit.
• _extract_intended_directory_from_pattern consults the first segment when
  given a comma-list (previously it tried the whole literal as a path).
• Optimizer's "no matches" warning now names the primitive instead of
  echoing the (potentially 80+ char) raw pattern, and still fires exactly
  once per primitive — not once per comma segment.

Design summary

Picked python-architect option (a): single shared helper instead of per-converter
splits, so the comma-split contract (trim, drop empties) is defined once. Kept
Instruction.apply_to: str because it is used as a dict key in
template_builder.py and a set member in claude_formatter.py; converting to
list[str] would have broken hashability and silently changed grouping.

Considered alternatives:

• Normalising applyTo once at primitive-load time in parser.py. Rejected:
  forces the type change above and changes serialisation everywhere; only the
  three converters and the optimizer actually need split form.
• Per-converter split. Rejected: triplicates the trimming / empty-segment
  rules and drifts.
• Emitting Cursor multi-glob as multiple [[rules]] blocks. Rejected: not
  required by Cursor's gray-matter parser — YAML list is accepted and far
  less invasive.

Validation evidence

Original repro from the issue: primitive with
applyTo: '**/src/**,**/api/**,**/services/**' compiled to all four targets:

===== copilot =====
---
applyTo: '**/src/**,**/api/**,**/services/**'
description: 'Rules for src, api, services'
---

# Multi-glob rules

Body content

===== cursor =====
---
description: Rules for src, api, services
globs:
  - "**/src/**"
  - "**/api/**"
  - "**/services/**"
---

# Multi-glob rules

Body content

===== windsurf =====
---
trigger: glob
globs:
  - "**/src/**"
  - "**/api/**"
  - "**/services/**"
---

# Multi-glob rules

Body content

===== claude =====
---
paths:
  - "**/src/**"
  - "**/api/**"
  - "**/services/**"
---

# Multi-glob rules

Body content

Optimizer on the same primitive (no misleading warning):

Placements: {PosixPath('/private/tmp/repro-1366'): [Instruction(name='multi', ...)]}
Warnings: []

Local verification gate:

• uv run --extra dev ruff check src/ tests/ — silent
• uv run --extra dev ruff format --check src/ tests/ — silent
• uv run --extra dev pytest tests/unit — 8762 passed (1 pre-existing
  unrelated failure: test_runtime_windows::test_execute_runtime_command_uses_shlex_on_unix
  — environment-specific codex path resolution, exists on main)
• New e2e test (tests/integration/test_apply_to_comma_e2e.py) — green
• New optimizer tests (TestApplyToCommaInOptimizer) — green
• New converter tests (TestApplyToCommaSplitting) — green
• New helper tests (tests/unit/utils/test_patterns.py) — green

Scope note

The Copilot target is intentionally untouched in this fix — it copies the
instruction verbatim, and the consuming tool handles the comma split. The
monolithic AGENTS.md / GEMINI.md compilers (agents_compiler.py,
distributed_compiler.py, claude_formatter.py for AGENTS injection) are
also out of scope; the issue specifically calls out the per-file Claude /
Cursor / Windsurf converters and the placement optimizer.

[Copilot]

Pull request overview

Fixes documented applyTo behavior by treating comma-separated values as multiple globs across the Claude/Cursor/Windsurf converters and the compilation context optimizer, while preserving Copilot’s verbatim behavior.

Changes:

• Introduces parse_apply_to() as a shared helper for trimming/splitting comma-separated applyTo values.
• Updates Claude/Cursor/Windsurf instruction converters to emit either a scalar (single glob) or a YAML list (multi-glob).
• Updates the context optimizer to treat comma-lists as “match any segment”, improves fallback directory inference, and reduces warning noise (names primitive instead of echoing the raw pattern).

Show a summary per file

File	Description
src/apm_cli/utils/patterns.py	Adds shared applyTo parsing helper used by converters/optimizer.
src/apm_cli/integration/instruction_integrator.py	Uses shared parser to emit correct multi-glob YAML for Claude/Cursor/Windsurf.
src/apm_cli/compilation/context_optimizer.py	Handles comma-lists in matching/placement and improves warning messages.
tests/unit/utils/test_patterns.py	Adds unit coverage for parse_apply_to().
tests/unit/integration/test_instruction_integrator.py	Adds converter coverage for comma-lists + Copilot verbatim regression test.
tests/unit/compilation/test_context_optimizer.py	Adds optimizer coverage for comma-lists and warning behavior.
tests/integration/test_apply_to_comma_e2e.py	Adds end-to-end test across all four targets.

Copilot's findings

• Files reviewed: 7/7 changed files
• Comments generated: 2