Read Mcode Skills with bounded concurrency

[SaladDay]

Mcode's first Turn can wait for serial inspection and stable reads of independent Skills. The pinned native patch reads up to four candidates per root concurrently, then commits results in candidate order. It preserves native file identity checks, cache keys, diagnostics, precedence and lifecycle ownership. The companion build runs the new native regression.

In a controlled local fixture with three selected Skills, three MCP bindings, real File/View and a fake provider, adding 100 ms to each File operation reduced first-Turn-to-first-provider-HTTP latency from 51.87–51.93 s to 40.14–40.18 s. Native-factory-to-first-HTTP fell from 67.08–67.18 s to 55.37–55.40 s, with the same 692 File calls and 597 Lookups. Zero-delay total was 3.368 vs 3.349 s, which establishes no meaningful gain. Selected Skill reads still took about 31.5 s at 100 ms. These measurements end at the local count_tokens request and do not represent live inference or full environment initialization.

Validation:

• Pinned-source Skill, prompt and MCP lifecycle regressions pass without skips; strict Skill TypeScript check passes.
• Seven baseline/candidate semantic comparisons cover cache reuse/change/pruning, precedence, symlink rejection, completion order and replacement between inspection and read. An additional native regression checks reversed candidate failure completion preserves diagnostic order.
• Package tests and syntax checks pass; three native binary qualification gates remain for the coordinated component build and live acceptance.
• Root review and fresh independent blind review found no P0–P2 issue; the final test-only addition received focused validation.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}