S0088-cli-deid-1: redact produces a real de-identified copy, delegated to @cosyte/deid

[NSchatz]

Spec: work/specs/S0088-cli-deid-1/spec.md (umbrella).

cosyte redact <file|-> (alias deid) was an honest, typed refusal gated on a ground
layer that now exists. This wires @cosyte/deid behind the existing seam
(src/core/deid.ts) as an optionalDependency, on the lazy per-format dynamic-import
pattern the breadth parsers already use. The CLI adds no policy, no locus map, no
transform and no fallback scrub: it delegates, or it refuses.

What it does now

For ccda, fhir, hl7, x12: stdout carries the de-identified document,
serialized exactly as cosyte fmt serializes that format. stderr carries the library's
own value-free manifest (one line per locus: category, transform, structural path,
count, disposition, stable code) plus the library's own published output label and
version. The CLI asserts no de-identification standard of its own.

What it refuses, and why the codes differ

Outcome	Diagnostic	Exit
every locus handled	the manifest	0
a locus the library could not handle	CLI_DEID_INCOMPLETE (new)	1
astm, mllp, ncpdp: no adapter there	CLI_NOT_IMPLEMENTED	69
dicom: covered there, binary result this text stdout cannot carry	CLI_FORMAT_UNSUPPORTED	65
@cosyte/deid absent from the install	CLI_PARSER_UNAVAILABLE	69

Every non-zero outcome emits nothing on stdout. The two refusals are deliberately
distinct: DICOM is the CLI's own channel limit, not a gap in the library, and blaming
the library for it would be a false attribution. @cosyte/deid/ncpdp is NCPDP
Telecom while this CLI resolves SCRIPT, so the subpath exists and the coverage
does not.

Two things worth a reviewer's attention

• The happy path was measured, not assumed. Under the default policy the existing
  adt-a01.hl7 and patient.fhir.json come back clean, but 834.edi drives ten
  blocked loci, so it is now the fail-closed fixture and two synthetic X12/C-CDA
  fixtures carry the clean path. No filter hides a blocked entry.
• The default policy needs a key. It pseudonymizes MRN / account / beneficiary
  numbers, and a keyed transform with no context is a fatal there, never an unkeyed
  fallback. The CLI holds no key material and adds no key flag, so it keys each
  invocation with an ephemeral random key and discloses on stderr and in the docs
  that surrogates are consistent within one output and not stable across runs.

Gates

typecheck, lint, format:check, phi-scan, check:no-emdash,
check:no-internal-refs, check:agent-notes, smoke and attw all green. 641 tests
(was 524); coverage core 97.8 / commands 98.2 against the 90 gate. The packed
tarball was installed from a directory outside the repo and both bins run.

No published exit value moved and no existing CLI_* code was renamed. redact stays
off the MCP tool surface, and it deliberately does not honour --unsafe-show-values.