[perf-scan] 2026-09-17 -- performance opportunities found

Performance Scan - 2026-09-17

Automated scan of src/apm_cli/ for algorithmic performance anti-patterns.
2 finding(s) identified.

Findings

[A] Quadratic loop nesting -- src/apm_cli/commands/uninstall/engine.py:189-230

• Current: O(P*S) where P = packages_to_remove and S = surviving_dependencies.
  _stage_shared_local_survivors iterates packages_to_remove (line 189) and
  calls _surviving_local_refs_at_install_path for each package, which itself
  scans the full surviving_dependencies list (engine.py:85-104) and
  re-parses every entry with _parse_dependency_entry(...).get_install_path(...).
  For an uninstall of many packages against a large surviving-dependency set
  this becomes quadratic re-parsing/re-resolving of the same install paths.
• Proposed: O(P+S). Build a dict[Path, list[DependencyReference]] once by
  parsing surviving_dependencies and grouping by get_install_path()
  before the packages_to_remove loop, then do an O(1) dict lookup per
  removed package instead of re-scanning and re-parsing the survivor list
  each time.
• Fix: Pre-index surviving_dependencies into a {install_path: [survivor_refs]}
  map once (parsing each survivor exactly once), then replace the call to
  _surviving_local_refs_at_install_path(package, surviving_dependencies, ...)
  with a dict lookup keyed by the removed package's own install path.

[C] Unconditional expensive operation recomputed per file -- src/apm_cli/install/manifest_reconcile.py:308-331

• Current: O(F*T) where F = deployed file paths and T = active+declared+known
  targets. reconcile_deployed_block's inner _target_for(path) closure
  (line 308) is invoked once per path via _locator(path) (called at lines
  327, 343, 353, 364), and on every invocation it re-iterates the full
  targets + declared_targets + scoped_known_targets list and calls
  install_governance([profile]) fresh for each profile (line 319) --
  recomputing the same per-target prefix/scheme sets for every single
  deployed file in the block.
• Proposed: O(F+T). Compute the per-target (prefixes, schemes) governance
  once per unique target name before the file loop (a dict[str, tuple[set,set]]
  keyed by profile.name), then have _target_for do O(1) dict lookups
  against that pre-built table instead of recomputing install_governance
  per path.
• Fix: Hoist a target_governance = {p.name: install_governance([p]) for p in ordered}
  dict above the _target_for closure (built once per reconcile_deployed_block
  call) and have _target_for iterate that pre-computed table instead of
  calling install_governance([profile]) inside the per-path loop.

Scan coverage

• src/apm_cli/ (484 files scanned)
• Patterns checked: A (quadratic loops), B (linear scan in loop),
  C (unconditional expensive ops), D (redundant config parsing),
  E (heavy top-level imports), F (sequential independent I/O)

Notes: Pattern E (heavy top-level imports) is already well-mitigated -- the
CLI's _LazyCommand mechanism in cli.py defers imports for install, uninstall,
pack, marketplace, prune, audit, and update to first dispatch. Pattern F
(sequential I/O) is already parallelized in registry/operations.py via
bounded ThreadPoolExecutor for MCP server checks/validation. No confirmed
findings for patterns B or D beyond noise (isolated dict/set membership
checks, not per-iteration linear scans).

Generated by Daily Performance Scanner · copilot · auto · 197.4 AIC · ⌖ 4.46 AIC · ⊞ 10.1K · ◷

• expires on Sep 24, 2026, 1:41 AM UTC

Start with _stage_shared_local_survivors in src/apm_cli/commands/uninstall/engine.py and reconcile_deployed_block in src/apm_cli/install/manifest_reconcile.py, tracing their current per-item scans and governance calculations. Verify that the optimization preserves existing uninstall and reconciliation behavior while removing repeated survivor parsing and per-file target governance work; the issue names no specific tests, so inspect nearby coverage before running it.