Promote target resolution to stable CLI and MCP

[jlitola]

resolve_target and githits resolve become default stable surfaces for fuzzy public package, repository and documentation-site names. The public MCP catalog contains 14 tools; Research remains experimental. MCP providers must now supply resolveTargetService, using the existing public /client implementation and /tools factory.

Resolver discovery now frames the agent task as identifying an OSS dependency’s canonical package, repository or docs target in a complete 79-character sentence. The descriptor and quick_start remove repeated argument, format and read-locator mechanics; the public MCP skill remains an exact guide copy. Compact result text gives uncertainty/empty-result instructions once, retains filter narrowing, and no longer claims multiple visible candidates for a singleton. Confidence/security gates, warnings, canonical locators, ranking, queries and JSON remain unchanged.

The implementation docs retain accepted launch limits, measured guidance/output sizes, neutral agent traces and the release/hosted rollout checklist. The pending fragment declares minor impact for both packages. Hosted clients receive the change after coordinated release, remote-mcp adoption and deployment.

Validation:

• bun test: 5,486 passing tests, zero failures across 233 files. Review corrections passed 159 affected adapter/guide/parity/smoke tests.
• Typecheck, lint, formatting, plugin generation/check and build passed. Existing packed-import/type/browser/artifact validation remains applicable; imports/exports are unchanged by the guidance refinement.
• Authenticated dev CLI/MCP smokes passed, including resolver text/verbose/JSON, selected-target inventory and JSON parity. Experimental Research checks also passed. Built Node CLI/MCP smokes passed without credentials.
• Exact text sizes: resolver description 1,121→580 bytes; full definition 3,342→2,799; guide 7,807→7,551 and public skill 8,338→8,082. Fixed ambiguous/empty cases 365→237 and 257→145; actionable and blocked fixtures unchanged. These are text sizes, not provider-token savings.
• Four follow-up Claude/Codex runs produced structured answers with zero recorded isolation violations. Claude used fuzzy resolution then source; both agents used site resolution/docs and preserved MEDIUM identity uncertainty. Both full-guidance runs omitted quick_start. Codex still used web with zero GitHits calls on the neutral fuzzy prompt, so that run does not prove resolver discovery. Whole-run metrics and adapter limitations are documented.

Luna promotion preflight and internal reviews passed. Claude Opus 5.5 promotion round 1 was clean after minor wording fixes; refinement round 2 was clean after restoring the ambiguous filter remedy; round 3 approved the user’s OSS dependency framing. Both included permitted fresh-context final checks. Both temporary plans were retired after their clean reviews. Final PR CI and MCP package validation passed at 1913f7e: build/checks, Linux/Windows tests, Bun and Node 20/22/24/26 compatibility. Publishing was skipped. Two new targeted Claude/Codex runs also used the resolver and preserved MEDIUM identity uncertainty, with no recorded isolation violations.