Deferred: `tan model` — NPU compile path

[alpCaner]

Deferred from the 0.5.0 RC (see the v0.5.0 milestone for the essential set).

tan model is ported and registered, but is not gap-closed against the Rust oracle and has no dedicated test coverage.

Why it defers cleanly: the host compiler it drives (dxcom for DEEPX) is license-gated, so no customer can exercise this path today regardless of how polished it is. Docs mention it 13 times, but mention is not use.

Scope: flag/exit-code parity against the released v0.4.1 binary, test coverage, and a decision on what it should do when the vendor toolchain is absent.

It must keep EXISTING and not error in 0.5.0 — shipped v0.4.1 binaries forward this verb, so a customer's muscle memory has to keep working.

[alpCaner]

Triage note — one claim in this issue is out of date. The issue stays open, narrowed.

Struck: "no dedicated test coverage".

python/tests/commands/test_model_command.py exists and collects 11 tests. It landed in 729234a (2026-07-30 15:45:15), before this issue was filed (2026-07-31T07:32:59Z).

tan model itself is a real implementation, not a forward — python/tan/cli.py:89, python/tan/commands/model_cmd.py (506 lines), which says so at model_cmd.py:17-22.

Still open, narrowed to two concrete gaps:

1. No byte-for-byte oracle diff for model. faultdecode and new-som both have one; model has none, and model does not appear in python/tests/parity/test_oracle_parity.py's CASES (lines 63-116) at all. "oracle" appears in test_model_command.py only in a docstring.
2. No dxcom-absent test. model_cmd.py:24-33 documents a deliberate generic model.build-failed divergence, but nothing asserts the toolchain-absence path.

Note on PR #58 (integration/model-edge-ai, "feat(model): tan model edge-AI lifecycle surface (check/zoo/add/prep/run/ab)"): it does not close this issue. Its six new subcommands have no v0.4.1 oracle counterpart, so they supply neither gap above — and they widen the surface this issue must diff. They also collide with #224 (new model.* codes with no Python emit-site gate to catch an unregistered one) and #261 (model declares zero of the seven global flags).