test(9p): run the conformance special-files cases over the wire

[pi0x]

The 9P half of #14: the conformance suite's five special files cases run
over the 9P column instead of skipping — 10 new cells in
.agents/conformance-matrix.md.

What moved

• test/9p/client.ts — p9Driver gains mountx.mknod. It is a Tmknod,
  with dev coming apart into major/minor the same 8-bit way
  P9Session.#mknod puts it back together. The client decides nothing: 9P2000.L
  carries the whole mode, type bits included, and the session hands it to the
  driver's extension unchanged, so EEXIST, ENOENT and EPERM-for-a-directory
  are the memory driver's own answers arriving as Rlerror.
• test/9p/conformance.test.ts — THROUGH_9P declares
  extensions: ["mknod"] (THROUGH_9P_REOPENED spreads it). The node-fs oracle
  target overrides it back to []: that driver implements no mountx.mknod and
  #mknod rightly answers ENOSYS there, which is the difference between a
  capability and a claim.
• test/matrix.ts, .agents/testing.md — the "only the loopback column has
  it" prose is replaced by which columns offer the extension by name and why each
  of the others does not.

What it buys over the session tests

test/9p/session.test.ts:1788 already proves the dispatch. These cases put a
FIFO through the rest of the surface over the wire: stat and readdir agreeing
on the seven type predicates decoded from Rgetattr's own mode, rdev surviving
the major/minor split and the single rdev[8] rejoining it, a special file
renamed and unlinked as an ordinary name, and each refusal arriving as this
transport's error reply.

On invariant 5

Declaring the extension here is not faking a capability: the client routes a real
Tmknod to a real session, resolveCapabilities would infer the same thing from
Object.keys(driver.mountx) anyway (src/harness.ts:44-45), and the columns
already declare capabilities as transport claims rather than driver truths. The
old "no wire representation" justification was the weaker half — Tmknod is one.
What keeps it honest is that the declaration names what that target delivers,
hence the oracle's override.

Not in scope

NFSv3 and NFSv4.1 keep the skip, and #14 now tracks why: both put the file type
in ftype3/nfs_ftype4 rather than in the mode, so two of the five cases cannot
route faithfully — NF4DIR is mkdir and NF4REG is NFS4ERR_BADTYPE — and
expressing partial carriage needs a suite change, not a wiring change. FUSE and S3
stay out for the reasons they always had.

pnpm test green (2985 passed, 310 skipped); pnpm matrix regenerated, the 9P
column moving from 192 passed / 18 skipped to 202 / 8.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
mountx	Ready	Preview	Jul 31, 2026 10:36pm