feat(codegen): wire [[codegen]].options through to tx3c, re-cut #128 regressions

[scarmuega]

Plan

plans/trix-codegen-options-wiring.md (tx3 domain root) — a
code-change-request scoped to this repository only.

Done criterion: met. Both parts land, the full suite is green, and each
ported regression was verified to fail when its fix is reverted.

What changed

Part 1 — wire the dead [[codegen]].options field

The field has been dead config since it was introduced: written None at
src/commands/codegen.rs and src/commands/init.rs, never read, and
src/spawn/tx3c.rs::codegen() forwarded only tii/template/output. A project
setting options got silence.

• src/config/convention.rs — KnownCodegenPlugin::default_options()
  and CodegenConfig::resolved_options(). Plugin defaults with the
  project's options merged on top, key by key, user values winning
  (including setting a default's key back to false). Plugin knowledge
  stays in trix's config layer; tx3c remains generic. The one default
  today is ts-client → {standalone: true} — the knob the
  codegen-v1beta0 templates gate package.json.hbs and
  tsconfig.json.hbs on. BTreeMap, so the JSON on the wire is stable.
• src/spawn/tx3c.rs — codegen() takes the resolved map and forwards
  it as one JSON object under --options. An empty map omits the flag
  entirely, so a project that asks for nothing produces exactly the argv
  trix sent before the channel existed.
• src/commands/codegen.rs — resolves once per [[codegen]] entry and
  passes it to every protocol in that entry.

Part 2 — PR #128's regression value, re-cut onto the #129 layout

#128 is built entirely on the tests/e2e/ layer #129 deleted and cannot
merge. Its three surviving regressions are re-homed at the layers the
current policy puts them:

• @→: forbidden-character rejection → tests/cli/refs.rs. Parse-time,
  zero registry traffic, so no stub is needed.
• image/png logo-layer pull (regression of fix(use): accept image/png logo layer when pulling protocol images #118) and lowercase OCI repo
  addressing with case-preserving identity (regression of fix(publish): lowercase OCI repository path for uppercase scopes #120) →
  tests/cli/registry.rs, against tests/harness/oci_stub.rs — test(e2e): regression coverage for ref grammar, logo-layer pull, OCI casing, package.json placement #128's
  in-process registry stub, ported. Pulls are trix's own code, not a helper
  binary, so they belong to the CLI layer and the suite stays offline.
  sha2 joins dev-dependencies: oci_client verifies every blob against
  its descriptor, so the stub's digests must be real.

#128's package.json fixture lock and #[ignore]d live test are deliberately
not ported — emission verification moves to the e2e journey, and the
trix layer asserts argv only (the #129 policy).

Closing #128 as superseded is the owner's act, not mine — flagging it
here per the plan.

Test placement note

The plan asks for the ts-client default to be asserted at the contract
layer. Reaching a known plugin's argv normally means fetching the web-sdk
templates from GitHub, which the hermeticity policy forbids. Both contract
tests that need ts-client use the existing local-template seam instead:
plugin.repo is treated as a local template root whenever it names a
directory, and repo resolves relative to the process CWD, so
stage_local_ts_client_templates() materializes
tx3-lang/web-sdk/.trix/client-lib/ inside the test project. No network, no
production seam added. The merge rule is additionally unit-tested next to
itself in src/config/convention.rs, per the "rules go innermost" doctrine.

tests/README.md gains a section on the OCI stub and the layering rule for
pull-path tests.

Verification

All run locally on macOS at 20eeebc, using CI's exact commands:

• cargo fmt --all -- --check — clean
• cargo clippy --all-targets --all-features -- -D warnings — clean
• cargo test --workspace --all-targets --all-features --locked —
  115 passed, 0 failed (83 unit, 13 CLI, 19 contract)

Regression spot-checks — each fix reverted locally in turn, only its own
test failing:

Reverted	Result
image/png dropped from oci::pull's accepted media types	use_accepts_image_with_png_logo_layer FAILED, other passed
scope/name no longer lowercased in oci::reference_for	use_addresses_registry_repo_path_in_lowercase FAILED, other passed
@ added to refs::validate_ident's charset	use_rejects_npm_style_at_version_separator FAILED

The tests/README.md invariant holds: no test in the repo spawns a real
helper binary (the only Command::new outside the harness's fake-tx3c
wiring is the rustc that compiles that fake).

Release ordering — please read before merging

--options is a flag tx3c does not have yet; its side is
plans/lang-tx3c-options-channel.md against tx3-lang/tx3, still ready.
The two are independent to develop (this repo's contract layer stubs tx3c)
but not independent to release:

• Projects with no [[codegen]].options and a plugin with no defaults are
  unaffected — trix omits the flag, so the argv is byte-identical to today.
• ts-client projects are not: they now always carry
  {"standalone":true}, so a released trix carrying this change paired with
  a released tx3c predating the options input would fail trix codegen
  for every TypeScript user.

The compat floor (src/spawn/compat.rs, tx3c min 0.22.0) is not
raised here — the version that carries the flag does not exist yet, and
raising the floor is the toolchain release train's step, not this plan's.
Filed as a follow-up draft plan in the domain root
(plans/trix-compat-floor-tx3c-options.md) for the owner to release. Merging
this PR is safe; shipping a trix release from it before the tx3c side is
released is not.

Not in this PR

Per the plan's boundary, the remaining parent-plan work stays where it is:
the web-sdk codegen-check.sh extension, the 07-codegen-consume journey
amendment, and the release-train note.

🤖 Generated with Claude Code