P5 row 106/107: extract busbar-plane-mcp to GetBusbar/busbar-plane-mcp (pinned); every gate reads the pinned planes as its base layer

[MattJackson]

P5 extraction, /status rows 106 (no busbar-plane-* crate under crates/) and 107 (crates down to 14). The mcp plane is the first of the five. Crates go 24 -> 23.

What

• crates/busbar-plane-mcp leaves for plane-mcp: busbar's in-tree crate, history imported, at busbar 00555bd98a (P5 extraction) busbar-plane-mcp#10: history imported with filter-repo; the logic crate is predev 00555bd's; busbar-plane-mcp-plugin (cdylib) runs busbar's published conformance suite.
  • The root pins it with one git row; the two edges in crates/busbar take the workspace row.
  • deny.toml allow-git gains the repo.
  • plugins.yaml drops pending_crate and sets busbar_checkout: true (an existing field), because its tests read busbar's battery and qa/method-inventory.json at the pin.
• The base layer (ARCHITECT ruling on row 106, option (A); spec deps: bump jsonschema from 0.49.2 to 0.49.5 #39 "plugin SOURCE never re-enters busbar"):
  • Ctx gains a BASE layer, beneath the overlay and above the disk. Ctx::workspace lays every pinned package of the plane kind at crates/<package>, read from the checkout cargo metadata resolves for the rev Cargo.lock pins (kind_isolation::plane_base). Nothing is written, copied or linked into the tree.
  • with_overlay keeps the base, so plants still win.
  • It is generic over the kind and names no plane.
  • It fails closed: a pinned plane that does not resolve is an Err naming the package and rev.
  • kind-isolation's own mount is untouched. No ledger row changes.
• Tests that read the filesystem resolve a pinned plane through the same rule (crates/busbar/tests/common/pinned_planes.rs):
  • capability-equality: the 5 proven mcp-server cells, and the >= 5 plane floor, which counts pinned members.
  • The kernel trust and core-admin plane-key ratchets.
  • xtask/tests/declared_classes.rs, now read through Ctx.
  • proto-deletion-gate.sh.
• The plane is judged where its source changes: plugin-ci.yml gains a plane step.
  • At the repo's busbar pin, it runs plane-purity, structure-lint and config-schema over busbar's pinned plane and again over the PR's logic crate (BUSBAR_PLANE_SOURCE).
  • A row that passes over the pin and fails over the change is RED (scripts/fleet/plane-gates.sh, with a selftest).
  • It takes effect in a plane repo once that repo pins a busbar that has it.

RED-able tests

• xtask/tests/pinned_planes_base.rs:
  • The base is laid from the checkout, and none of it is on disk.
  • A plant wins over the base.
  • structure-lint, config-schema, no-float-money, plane-purity and reachability each lose a passing row when the base is removed.
• declared_classes: the scan reads the pinned plane, and reads none of it without the base.
• capability_equality: the pinned plane's cited evidence is not on disk, and is real when read from the checkout.

Ledgers: none touched (ruling condition 6).

Not in this PR:

• The three row-159 HIGHs in busbar-plane-mcp (their fixes land in the plugin repo, then a pin bump here).
• H2's kernel caller_ref.
• crates/busbar-plane-decisions/tests/invariance.rs still derives its sibling planes from crates/ alone, so it no longer sees mcp. That file is held by decisions plane: route by model, splice upstream_model, one billing rule, token classes, no unit leak or eviction, zero-alloc relay (Q128 findings, rows 128/159) #806/decisions plane: the nanojev dialect (Q128 finding 3, row 128) #807.