refactor(availability): launch gate on the shared probe; typed IAdapterRegistry surface (#435, parts 3-4)

[debugmcpdev]

Parts 3–4 of #435, completing the issue (part 1: #437; part 2: #442).

Part 3 — launch gate on the shared probe

checkLaunchToolchain (the gate behind create_debug_session and start_debugging, issue #360) was the last hand-rolled getFactory→validate→fail-open path. It is now a thin wrapper over probeLanguageEntry with a synthetic entry — installed: true plus an empty disabledSet provably neutralize the probe's notInstalled/disabled short-circuits, so the gate keeps gating on the toolchain only (disabled languages are already refused upstream at create_debug_session). The gate's availability answer and reason text are now literally computeModeAvailability's launch mode, so it can no longer drift from list_supported_languages the way the server/doctor loops used to before part 1.

The full fail-open contract was pinned with direct unit tests before the rewrite (missing registry, throwing getFactory, missing validate, throwing validate, load errors, the empty-errors fallback sentence, one-validate-per-cache-window), and a drift-fence matrix asserts checkLaunchToolchain(...) ≡ probeLanguageEntry(synthetic).modes.launch across every registry shape — the gate-side analogue of the doctor parity test. The existing launch-gate and #360 carve-out suites pass unmodified.

Part 4 — typed registry surface + loader-error surfacing

IAdapterRegistry gains the members the server was reaching through as unknown as duck-typing: required listLanguages / listAvailableAdapters / getFactory / getFactoryMetadata (plus register aligned to its actual Promise<void>), and optional getFactoryResult. All four cast sites are gone (server.ts ×3, session-manager-operations.ts ×1; runtime typeof guards stay for partial test doubles) — a rename on the concrete AdapterRegistry now breaks the build instead of silently degrading every language to fail-open. Doctor's duplicate DoctorRegistry structural type becomes a Pick of the interface.

getFactoryResult closes the swallowed-loader-error hole: AdapterRegistry.getFactory's catch { return undefined } collapsed a corrupt package, a broken export, and dynamic-loading-disabled into one undefined, so probe.factoryLoadError could never fire. Now getFactory delegates to getFactoryResult (fail-open contract unchanged, one code path), the availability probe prefers getFactoryResult when the registry offers it, and doctor reports the loader's real error message instead of "(the registry returned no factory)". The two remaining discovery-fallback swallows log warnings. New shared types: AdapterManifestEntry (the loader-manifest shape, finally disambiguated from the factory-declared AdapterMetadata) and FactoryLoadResult.

Test fixtures: the central IAdapterRegistry mock provides fail-open-neutral defaults for the required members (mirroring the pre-typed fallback paths); the WithErrors/WithLanguages builders override the discovery members consistently. AdapterRegistry.getFactory/getFactoryResult get their first direct unit tests (registered / cached / loader-throws / dynamic-disabled).

Deferred (noted on #435 at close)

Consolidating the three near-identical doctor fake-registry builders; doctor wording for dynamicLoadingDisabled (production always enables dynamic loading); switching the #360 carve-out site to getFactoryMetadata; extending the isAdapterRegistry runtime guard to the new members.

Verified locally: full build, lint, 222-file/4048-test unit suite, and a real mcp-debugger doctor --json run (the real registry now feeds getFactoryResult through the shared probe).

Closes #435.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 98.14815% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/container/dependencies.ts	0.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!