feat(ext-proc): make external processor crate publishable

[nerdalert]

Summary

This PR makes the generic praxis-proxy-ext-proc crate publishable.

Currently ext_proc works inside the Praxis workspace, but downstream repos cannot depend on it through the normal released-crate path. That blocks llm-d integration test work in praxis-proxy/ai, where the AI gateway needs to register the same generic ext_proc filter without local-path or git-dependency workarounds.

ext_proc is not full Envoy parity yet, but publishing the crate unblocks the llm-d integration-testing epic.

What Changed

• Marks praxis-proxy-ext-proc as publishable.
• Exports register_filters() so downstream binaries can explicitly register the filter.
• Exposes praxis_ext_proc::proto so downstream code can use the generated Envoy ext_proc protobuf and gRPC types.
• Adds praxis-proxy-ext-proc to the Makefile publish dry-run crate list.

Generated Proto Lint Suppression

This PR exposes praxis_ext_proc::proto so downstream crates can use the same generated Envoy ext_proc protobuf and gRPC types as Praxis. That is needed for downstream integration tests and external processor implementations that need ExternalProcessor, ExternalProcessorServer, ProcessingRequest, and ProcessingResponse.

This is a new public API shape for Praxis filter crates: most published filters expose handwritten Rust filters/config types, while praxis-proxy-ext-proc also exposes generated tonic/prost code.

Because the tonic-generated gRPC client/server methods are generated code, they do not include Praxis-style rustdoc sections such as # Errors on every public Result returning method. The PR keeps Praxis lint standards for handwritten code and suppresses generated-code-only Clippy lints only at the include_proto! module boundary.

Validation

• git diff --check
• cargo +nightly-2026-03-28 fmt --all -- --check
• make lint
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test -p praxis-proxy-ext-proc
• cargo test -p praxis-proxy-ext-proc -- --test-threads=1
• cargo deny check
• cargo doc --workspace --no-deps --document-private-items
• cargo package -p praxis-proxy-ext-proc --allow-dirty --no-verify

Note: full cargo package verification depends on release ordering and will pass after the next praxis-proxy-filter crate version is published first.

[praxis-bot]

ext-proc publish review

Clean, minimal PR. The visibility changes (pub(crate) to pub for proto and its inner modules), export_filters! usage, clippy::missing_errors_doc lint suppression for generated gRPC code, publish-crate ordering in the Makefile, and doc updates all look correct.

Severity	Count
Critical	0
Large	0
Medium	1

Medium -- no test for new register_filters() API: see inline comment.

[praxis-bot]

[Medium] The export_filters! invocation generates a new public register_filters() function -- the recommended API documented in the crate doc comment -- but no test in this crate exercises it. The existing pipeline_builds_with_ext_proc_and_failure_mode test manually registers via registry.register("ext_proc", http_builtin(...)) instead.

Update that test to use register_filters() so the published API has end-to-end coverage:

let mut registry = praxis_filter::FilterRegistry::with_builtins();
register_filters(&mut registry);

[shaneutt]

I would prefer not to publish this crate. I want to aim for it to not exist at all in time.

I suggest we transfer this crate over the the ai repository, and wall it off for everything except llm-d, and note that it's only for llm-d, keep the anti-pattern warning language, and keep it as an experimental feature.

We should explicitly claim no support or backwards compatibility for any use case except llm-d.

We run the risk of people building on it for other use cases and then it becomes quite hard to remove it later. Bare minimum, I want any such usage outside of llm-d to be gated behind a github request to discuss the use cases beyond llm-d.

[nerdalert]

Superseded by praxis-proxy/ai#334