test: fuzz targets for query binder, error writer, roster enum parse

[SyniRon]

What

Coverage-guided Go fuzz targets for the three panic-prone pure-function seams PRD #112 flags as the lenient request surface:

• FuzzQueryBinder (rest) — no accessor panics on any url.Values; first-failure-wins is monotone (a later successful read never clears an earlier 400, the invariant that makes the "bind everything, check Err() once" handler protocol safe).
• FuzzWriteError (rest) — arbitrary user-derived message bytes (control chars, invalid UTF-8) always yield valid gRPC-status contract JSON: code echoed, details: [], code→HTTP-status mapping intact, Content-Type: application/json.
• FuzzParseRosterType (types) — never panics on the base-0 name-or-number enum forms (hex/octal/binary/underscore/overflow/negative/empty); any clean parse round-trips through its enum name.

Why

These exercise the binding/error seams at millions of cases/sec in-process — coverage the golden corpus (recorded inputs) and Schemathesis (network edge) can't reach. They surfaced during a live smoke of the new stack; the network side came back clean, and these lock the unit seams down permanently.

Cost

Standard testing.F targets: a normal go test ./... runs only their seed corpora as fast deterministic unit tests — no fuzzing cost in CI unless -fuzz is invoked explicitly. CI can optionally add a bounded -fuzztime pass.

Verification

~870k execs across the three during the smoke, zero crashes, zero invariant violations. Seed corpora green on a clean develop base.

This was generated by AI