feat(pxq): Stage C — schema loading + pxf_* extensions + pxf_proto

[trendvidia]

Summary

Schema-aware query layer for the jq-style tool. Closes the README's
extension surface: all five pxf_* functions are live, and
-p schema.proto is wired to a descriptor registry that
pxf_directive("dataset") (for row binding) and pxf_proto(...) (for
typed output construction) consume.

Strict-mode compile-time AST validation against the schema stays
deferred to a follow-up — Stage C is "schema available, errors at
runtime" rather than "schema enforced at compile time."

Function surface

Function	Purpose
pxf_directive(name)	Directive list by name, source order. Recognised names: "type", "dataset", "proto", or any generic @<name>. Dataset rows are schema-bound field-name maps when a descriptor is in scope; otherwise schemaless. Proto entries surface shape / typeName / body.
pxf_fieldnames	Declared field names for the bound message; falls back to the document's @type when input is the top-level body map. Errors in loose mode.
pxf_type	Proto type of the input value, or jq-level type as fallback.
pxf_has(field)	Schema-aware has() that distinguishes absent from zero-value.
pxf_proto(name; obj)	Typed object construction. Validates the descriptor exists, attaches an @type sentinel the emitter lifts into a document-level @type Name directive.

Naming convention

gojq reserves @<name> for built-in string formatters (@uri, @base64, …) and doesn't allow third-party @-extensions. The README's @pxf.* surface lands as pxf_* plain gojq functions — same namespacing intent, valid grammar. README updated in this PR.

Architectural changes

• loadedDoc shape: adapters previously injected __pxf_datasets / __pxf_protos / __pxf_directives synthetic keys at the top level of the gojq graph (Stage A scaffolding). Stage C separates body from directives; pxf_directive(name) is now the documented access path. CSV's synthetic dataset wrapper becomes a real pxf.DatasetDirective so it round-trips identically with native PXF datasets.
• Emitter: an @type map key on the result is lifted into the document-level @type Name directive form on output, so pxf_proto(...) produces valid PXF on the wire rather than the invalid @type = "Name" field-assignment form.

Demo

The README's quick-start example now works as written:

$ pxq 'pxf_directive("dataset")[0].rows | map(select(.symbol == "AAPL")) | length' trades.pxf
value = 2

Test plan

• go build ./... clean
• go test ./cmd/pxq — 45 tests pass (28 from A/B + 17 new)
  • Schema loader: nested messages registered, compile errors
    propagate, empty-args returns nil
  • pxf_directive: type/dataset/proto/generic/missing cases
  • pxf_fieldnames: no-schema error, schema-resolved field list
  • pxf_type: jq-level fallbacks for all six kinds
  • pxf_has: absent vs present cell semantic
  • pxf_proto: no-schema error, successful bind with @type
    emission, unregistered-type error, pipe-form invocation
• README quick-start smoke test verified end-to-end
• CI passes on the matrix

Follow-ups (queued)

• In-document @proto directive as a schema source (Stage C2)
• protoregistry resolution (-s/-n/--schema)
• Strict-mode compile-time AST validation
• pxq infer-schema subcommand