fix: skip Form B packages in template name resolution

[xv1rcn]

Summary

• fetch_template_package() now skips Form B packages whose manifest
  is synthesized from the descriptor (mcpp = { ... })
• Uses existing mcpp::manifest::extract_mcpp_field() API, consistent
  with prepare.cppm Form A/B dispatch
• 4 lines added in src/scaffold/create.cppm

Refs #265 (hardening — does not close it; see review comment below for the
verified root cause, which lands in a separate PR)

Test plan

• bash tests/e2e/69_package_templates.sh — new Form B filter case
• bash tests/e2e/02_new_build_run.sh — no regression
• bash tests/e2e/12_add_command.sh — no regression
• bash tests/e2e/27_namespace_dependencies.sh — no regression
• CI: Linux / macOS / Windows

[Sunrisepeak]

Thanks — reviewed and reproduced locally. The Form B filter is worth having, but it is not what makes #265 fail, so I changed Closes #265 → Refs #265.

What actually flips the resolution

Not "imgui.lua unavailable for some reason" — the index floor gate itself. check_index_floor() conflates report once with enforce (src/pm/index_contract.cppm:110), while its only caller uses the return value as the gate (src/pm/package_fetcher.cppm:593). So within one process the first lookup into a floor-violating index is blocked and every later one is silently let through.

fetch_template_package() looks up twice — ns="" then ns="compat":

lookup	candidates	result
ns=""	imgui.lua, compat.imgui.lua	blocked by floor (E0006 printed once) — imgui.lua never seen
ns="compat"	compat.imgui.lua	gate already "reported" → passes → Form B hit

Reproduced with the exact binary CI installs (0.0.100) against the real index (floor 0.0.102):

$ .../xim-x-mcpp/0.0.100/mcpp new abc1 --template imgui
error: index requires mcpp >= 0.0.102 but this is mcpp 0.0.100 [E0006]
error: package 'imgui@1.92.8' has no mcpp.toml        # byte-identical to CI

$ MCPP_INDEX_FLOOR=ignore .../0.0.100/mcpp new abc2 --template imgui
     Created abc2 (template imgui@0.0.6:window)       # same binary, same index

Same binary, same index, only the gate differs — so candidate ordering and Form A/B are not the cause. The leak also defeats the floor contract everywhere else (prepare.cppm uses the same gate), which is exactly what mcpp-index #107 raised the floor to prevent. Introduced in 53b6ead (0.0.85, #200), i.e. present since the feature landed.

With this PR applied the same CI step still exits 1 — just with template package 'imgui' not found in the index instead. A separate PR fixes the gate and will close #265.

On merging this

Fine to merge as hardening. Verified it cannot regress anything today: imgui.lua and templates.lua have no mcpp field at all (Absent), all 40 TableBody descriptors in the index are compat.* source packages, and a TableBody package could never have passed the root/mcpp.toml check anyway — so the skip only turns an unavoidable failure into a clearer message. It is also textually orthogonal to the gate fix.

Two things before merge:

1. Status checks have not run on this fork PR — please let Linux / macOS / Windows CI go green first (the last box in your test plan).
2. Consider a one-line comment at the continue stating why Form B can never provide templates, so the skip is not mistaken for the mcpp new --template imgui fails on all platforms: package 'imgui@1.92.8' has no mcpp.toml #265 fix later.