Frozen license labels make every unrecognized string fail the release gate with HTTP 422

The piece of this design I would keep is the default branch in `classify`. A missing license value becomes an empty string. Anything that is not a key in the table comes back as class `unknown` with gate `fail_closed` ^[9]. The only route to `allow` is an explicit row, and the sample table has four ^[8]. An agent drafting the handler cannot produce a permissive verdict by leaving something out. It would have to add a row. The author put the rules in a table "so a reviewer can argue with the policy instead of arguing with generated branches" ^[16].

The author wrote that "The useful outcome is a failing test for unknown licenses, not a longer handler that merely sounds complete." ^[17] The sample meets that bar on its first request. Its snapshot lists `widget` with an empty license string ^[12]. Empty is not a key, so `widget` classifies as unknown, `blocked` goes true, and `GET /inventory` returns 422 ^[1].

The design breaks at string matching. The lookup is an exact key match after a whitespace strip ^[9]. `mit`, `Apache 2.0`, `GPL-3.0-or-later` and the compound expression `MIT OR Apache-2.0` all miss the four keys and fail the build ^[2]. The author says the table is not a complete SPDX catalog, and that readers should swap in obligation labels their own counsel has already approved ^[7]. For a release gate I think this is the right direction to fail. A false block costs a reviewer some time. A false allow is what the author calls a product bug, a guess shipped as an authoritative decision ^[3]. The price is a table someone has to own, plus a normalization step if your manifests emit free-text license names.

The status code collapses two outcomes. `GPL-3.0-only` maps to `block` ^[8]. `build_report` treats `block` and `fail_closed` the same way ^[10]. So a known strong-copyleft package and a string nobody recognizes both return 422 ^[3]. A CI step that checks only the status cannot tell them apart. The per-item `gate` field can ^[3]. I would have the gate consumer read `items`.

The review page is the best engineering in the piece. Four checks go in writing before any generated file is accepted. Every snapshot package appears exactly once, empty strings included. An unrecognized string never gets `permissive` or `allow`. Any `fail_closed` means `blocked` and 422. And a generated diff may not delete the pure-function tests to get the suite passing ^[13]. That last check is aimed at agents; a suite made green by deleting its tests still reports green. Keeping the classifier in a pure function means the HTTP layer cannot hide a bad label, and the server uses only the Python standard library ^[15].

The author labels the code a proposed workflow, not a log of a production deployment ^[14]. The claim that agents often guess at missing or ambiguous license strings comes without a count ^[2]. For the pattern to transfer, your dependency tooling has to emit strings that match your table's keys. Otherwise formatting noise sets the unknown rate. The checklist also asks only about direct dependencies before a build is tagged ^[18].