perf(adapters): metadata-only availability probe — stop importing all 9 adapters on tools/list

[debugmcpdev]

Fixes #401

On the dist/source distribution, adapters load lazily per session — but the availability probe defeated that: isAdapterAvailable was literally try { await loadAdapter() }, so the first tools/list (and create_debug_session / list_supported_languages via the same getSupportedLanguagesAsync() path) imported and instantiated all nine @debugmcp/adapter-* packages, permanently, into the heap of a server that may only ever debug Python.

Change

• New injectable PackageResolver seam on AdapterLoader (third constructor arg, mirroring the existing ModuleLoader seam). createDefaultPackageResolver():
  1. createRequire(import.meta.url).resolve(packageName) — presence via Node resolution, no module execution. ERR_PACKAGE_PATH_NOT_EXPORTED counts as installed (package present, ESM-only exports map).
  2. Fallback: fs.access over the loader's two existing monorepo fallback paths.
• isAdapterAvailable: cache-hit short-circuit (a genuinely-loaded factory) → resolver probe. No import, no factory instantiation.
• loadAdapter is untouched — full import + instantiation happen at the first real use of a language (AdapterRegistry.create()).
• listAvailableAdapters metadata comes from its static known-adapter table as before; the live-factory attach refinement still applies when a factory was actually loaded.

Scope decisions (as discussed in the issue)

• list_supported_languages' per-language getFactory() for modes + toolchain validate is deliberately unchanged: it's an explicit informational call with a 30s validation cache, and skipping the factory would silently turn launch.available from a real toolchain probe into assumed-true. The issue's stated goal — bounding "the server that answered initialize and one tools/list" — is met by fixing the probe.
• Trade-off: a present-but-broken package (missing factory export, corrupt build) now reports installed: true and surfaces its error at first load, with the existing install-hint message.
• Container mode unaffected (adapters pre-registered into the registry); npx bundle unaffected (batteries-included imports all 9 eagerly by design).

Measured (mem-bench, this machine, Windows)

node scripts/mem-bench.mjs --target dist --scenario idle,tools, median RSS over 5 trials. The metric for this issue is the tools/list step cost (after-tools-list − after-initialize):

	after-initialize	after-tools-list	tools/list step
before	109.2 MB	114.1 MB	+4.9 MB
after	98.1 MB	98.8 MB	+0.7 MB

(The absolute after-initialize difference is #400's lazy transports, which this branch is rebased on; the before-baseline predates it.)

Tests

• New metadata-only availability probe describe: probe returns availability without any moduleLoader.load call; cache short-circuit skips re-probing; listAvailableAdapters covers all 9 with zero imports.
• New createDefaultPackageResolver describe: resolve-success, exports-map-rejection-as-installed, fs fallback (platform-safe file URLs), nothing-resolves.
• The old load-based availability tests were superseded and removed — two of them had also silently become dependent on the local packages/*/dist being built (real fs hits).
• Full npm test green, npm run lint clean.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!