chore: normalize xpkg identity layout, enforce it in lint, pin CI to mcpp 0.0.105

[Sunrisepeak]

TL;DR

The index was already clean on identity — all 49 descriptors declare a fully-qualified name with an atomic short segment, so mcpp#278's split form appears nowhere here. No namespace or name value changes in this PR.

What was not pinned was the layer below: where a descriptor lives. That gap had let two real defects in.

Normalization

Change	Why
pkgs/c/capi.lua.lua deleted	Byte-identical duplicate (same md5) of pkgs/m/mcpplibs.capi.lua.lua — same identity (mcpplibs.capi, lua). Whichever the directory scan reached first won; edits to the other were dead code nobody would notice.
pkgs/t/tensorvia-cpu.lua → pkgs/a/aimol.tensorvia-cpu.lua	Canonical path for identity (aimol, tensorvia-cpu). Off-canonical names still resolve, but only through a COMPAT filename fallback mcpp schedules for removal in 1.0.0.

Lint now enforces all three identity rules

Rule 1 — name == <namespace>.<short> (existing, kept).

Rule 2 — short is a SINGLE atomic segment; depth belongs in namespace. (new)

namespace = "mcpplibs.capi",  name = "mcpplibs.capi.lua"   -- ✅ (mcpplibs.capi, lua)
namespace = "mcpplibs",       name = "mcpplibs.capi.lua"   -- ❌ short = "capi.lua"

Both survive mcpp's normalizer, because it splits the FQN on its last dot — so the second silently resolves to (mcpplibs.capi, lua), a namespace the descriptor never declared. After this rule the declared namespace and the canonical one are always the same string.

Rule 3 — descriptor sits at the canonical path for its identity. (new, tests/check_package_filename.lua) — this is the check that found both defects above.

(mcpplibs, cmdline)       → pkgs/c/cmdline.lua
(compat,   zlib)          → pkgs/c/compat.zlib.lua
(aimol,    tensorvia-cpu) → pkgs/a/aimol.tensorvia-cpu.lua
(mcpplibs.capi, lua)      → pkgs/m/mcpplibs.capi.lua.lua

The filename is not part of identity (mcpp resolves identity-first and re-verifies every hit against the file's own package.{namespace,name}), so this is about two other things: keeping a package on mcpp's first filename probe, and making duplicate descriptors for one identity impossible to add unnoticed.

Zero-namespace packages are untouched

imgui / ffmpeg / opencv keep namespace = "" — their bare name is the FQN, which is a deliberate, legal form. Consumers keep writing imgui = "0.0.3"; xlings key and store dir unchanged. Rule 2 still guards them: a dotted name with no declared namespace would resolve into a namespace written down nowhere.

CI → mcpp 0.0.105

env.MCPP_VERSION and all three matrix entries bumped 0.0.102 → 0.0.105, where mcpp xpkg parse enforces Rule 1 itself (INV-NAME, mcpp-community/mcpp#278). The two lua lints therefore become an earlier, cheaper second gate rather than the only defense.

index.toml's min_mcpp/latest_mcpp deliberately stay at 0.0.102. The floor is a consumer contract; 0.0.105 carries a breaking change (bare names no longer reach third-party namespaces) and nothing in this index needs 0.0.105 grammar — so there is no reason to force every consumer forward. Rationale recorded inline in validate.yml.

Verification

Run locally against real mcpp 0.0.105:

• All 48 descriptors pass every lint (syntax, required fields, leading-v, mirror URLs, name, filename) plus mcpp xpkg parse.
• Three counterexamples each rejected with the exact spelling to write instead:
  • split form (ns="chriskohlhoff", name="asio")
  • non-atomic short (ns="mcpplibs", name="mcpplibs.capi.lua")
  • dotted name with no namespace (ns="", name="mcpplibs.capi.lua")
• Both pre-normalization files are caught by Rule 3 when replayed through the new lint.
• Neither renamed package is a workspace member, so no test-example churn.

Docs

docs/repository-and-schema.md gains a 包身份:(namespace, name) 的三条规则 section (rules + examples + why filenames are pinned even though they aren't identity), and the repo-layout / CI-behaviour / local-lint-reproduction sections are updated to match.