feat(sled): state, local and education procurement

[makegov-mark]

SDK support for Tango's new /api/sled/ namespace — state, local and education (SLED) procurement. Blocked: the endpoints are not live on the public API yet. Hold until the API ships them.

Siblings: makegov/tango-python#62, makegov/tango-node#18, makegov/tango-go#6.

What's here

crates/tango/src/resources/sled.rs, plus five constants in shapes.rs:

Method	Endpoint
list_sled_opportunities / iterate_sled_opportunities	GET /api/sled/opportunities/
get_sled_opportunity	GET /api/sled/opportunities/{opportunity_id}/
list_sled_opportunity_revisions	GET /api/sled/opportunities/{opportunity_id}/revisions/
get_sled_coverage	GET /api/sled/opportunities/coverage/
list_sled_forecasts / iterate_sled_forecasts	GET /api/sled/forecasts/
get_sled_forecast	GET /api/sled/forecasts/{forecast_id}/

Four bon-derived options builders name every one of the API's 27 solicitation filters and 13 forecast filters.

Four things a reviewer should check

The SDK must not synthesize status=open. The API defaults a list request with no status and no active to open-only. Making that explicit client-side breaks active = Some(false), which is the complement of open and therefore has to reach the server as the only liveness param. no_liveness_filter_is_sent_when_none_requested and active_false_is_sent_not_dropped pin both halves.

Option<bool>, not bool, for the three boolean filters. active, has_documents and source_declared are optional so false reaches the server instead of collapsing into the Default — the same reason push_opt_bool exists in internal.rs. Getting this wrong would silently make "closed solicitations" and "records with no documents" unqueryable, with no error to notice.

SHAPE_SLED_REVISIONS_MINIMAL omits changes. The per-field before/after requires a Small plan; naming it in a suggested shape would 403 every Free caller on a field they never asked to gate. changed_fields is in the shape and available at every plan. A test asserts the split rather than trusting the constant to stay right.

change_seen_after has no _before twin, deliberately. The API exposes only the lower bound on that field. The doc comment says so, rather than leaving a reader to assume the pair was forgotten.

Scope note

This SDK has no dibbs, exclusions or sbir resources, so — as with tango-go — there was no vendored contract to refresh and no conformance or shape-coverage gate to satisfy. SLED lands as a plain resource module following the protests.rs shape, returning Record / Page<Record> like every other list resource here.

Testing

• cargo test --workspace — 191 unit tests pass, 16 of them new (plus the existing 16 + 19 + 6 + 4 across the other suites).
• cargo clippy --workspace --all-features -- -D warnings — clean.
• cargo fmt --all -- --check — clean.

Docs

New State & Local — SLED section in docs/API_REFERENCE.md, and an addition to docs/WEBHOOKS.md's alert semantics: an exclusion or DIBBS solicitation reaching its date fires nothing, because open/closed is derived at query time — but alerts.sled_opportunity.match does fire when a SLED solicitation closes, since its liveness is a stored field the API recomputes every fifteen minutes.

🤖 Generated with Claude Code

---

Update: the paid document-body leaf

Pushed a second commit adding attachments(extracted_text) — a SLED document body, served on the Small plan and above, from Tango API 4.25.1.

Three rules are documented wherever a caller will look, because each one is otherwise a support ticket:

• The leaf must be named. No default or suggested shape includes it, and attachments(*) does not carry it — the API resolves the body only for a caller who asked, so a default naming it would make every detail fetch pay for a document nobody wanted to read. A test pins the shapes against it.
• The key is absent, not null, whenever the text is not being served: below Small (where it is withheld and named in meta.upgrade_hints), on a contested document, or where the text could not be resolved.
• A contested document never returns text, at any plan — its stored bytes disagree with what the record advertised, so its extracted text is not reliably that record's content.

And the distinction worth stating plainly, because it reads backwards in both directions: searching document text is ungated (every plan, ids only, no fragment of the document) while reading a body is paid (Small and above, one record at a time). Buying the body does not change search.

Sequencing: the SLED surface itself is live on the production API. extracted_text needs 4.25.1, which is not deployed yet — so this PR is mergeable, but that one field will 400 until the API catches up. GET /api/version reports what a caller is talking to.