perf(compile): fix quadratic and redundant-resolve costs in placement

[Sergei Kozyrev (mukkumayc)]

Motivation

I hit a real performance problem running apm compile on a large internal monorepo, roughly 159,000 directories and 294,000 files, with several scoped applyTo patterns. Compile time was bad enough that I dug into ContextOptimizer and found the causes below.

I mostly write JS, not Python, so I used Claude Code to implement this fix. I reviewed the diff and verified it with a before after benchmark, but flagging this for transparency.

Summary

ContextOptimizer had several costs that only show up once a project has multiple scoped (non-universal) applyTo patterns, each matching a large, overlapping set of directories, the shape of a real monorepo with per-module conventions. Four issues stood out:

• _file_matches_pattern and _safe_recursive_glob resolved every file path independently per pattern instead of reusing a memoized relative path.
• _calculate_distribution_score recomputed the depth mean inside its variance loop (O(n²)) and rescanned _directory_cache for a count that's just its length.
• _is_hierarchically_covered, _find_minimal_coverage_placement, and _is_instruction_relevant resolved paths that were already canonical _directory_cache keys.
• _optimize_single_point_placement rescanned every matching directory per candidate instead of short-circuiting on the first miss.

No behavioral changes. Only removing provably redundant work.

Benchmark

Before/after comparison on synthetic monorepo-shaped projects: https://github.com/mukkumayc/apm-placement-benchmark

directories	before	after	speedup
~6	0.28s	0.27s	1.0x
~1000	2.90s	0.58s	5.0x
~2000	8.39s	0.92s	9.1x
~10000	158.00s	3.61s	43.8x

Test plan

• Full existing unit/integration test suite passes unchanged

[Sergei Kozyrev (mukkumayc)]

@microsoft-github-policy-service agree

[Copilot]

🟡 Changes recommended

The newly introduced memoization caches can grow unbounded across repeated compiles because they are not cleared when the long-lived ContextOptimizer instance is reused.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Improves apm compile placement performance by removing redundant path normalization and repeated scans inside ContextOptimizer, targeting large monorepos with multiple overlapping scoped applyTo patterns.

Changes:

• Memoizes per-file relative-path strings and uses cached Set[str] membership for ** glob matching.
• Fixes an O(n^2) mean/variance recomputation in distribution scoring and removes redundant directory-cache rescans.
• Short-circuits hierarchical coverage checks to avoid full rescans per candidate placement.

File summaries

File	Description
src/apm_cli/compilation/context_optimizer.py	Adds memoized relpath + resolved-dir caches and refactors matching/coverage calculations to reduce redundant work and quadratic behavior in placement optimization.

Review details

Suppressed comments (1)

src/apm_cli/compilation/context_optimizer.py:168

• _resolved_dir_cache is introduced as an unbounded memoization dict but is not cleared between compile runs. Since ContextOptimizer is reused across compiles, this cache can grow without bound and retain resolved Paths long-term; it should be cleared when starting a new placement optimization (similar to _glob_cache).

        self._inheritance_cache: builtins.dict[Path, builtins.list[Path]] = {}  # (#171)
        self._resolved_dir_cache: builtins.dict[Path, Path] = {}

• Files reviewed: 1/1 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

🟢 Approval recommended

The changes appear to be performance-only refactors with cache invalidation handled per run, and the only identified issue is a small maintainability cleanup (dead try/except) with a straightforward fix.

Review details

• Files reviewed: 1/1 changed files
• Comments generated: 1
• Review effort level: Lite