fix(generator): compile 20 real-world specs, gate CI on the working set

[lightsofapollo]

Summary

Ran scripts/spec-compile.sh against the 50+ OpenAPI specs in specs/ to surface generator bugs across real-world documents. 20 of 54 specs (gitea is Swagger 2.0 — skipped) now compile cleanly via cargo check. The remaining 34 surface other generator bugs that need deeper work; this PR documents them and gates CI on the working set so we don't regress what works.

Bugs fixed (in order of impact)

1. r#self panic — self/super/crate/Self cannot be raw identifiers in Rust. proc_macro2 panics outright when fed r#self. Specs that name a field self (datadog-v2, github, microsoft-graph, snyk, google-calendar, digitalocean, launchdarkly, …) all hit this. Fix: rename to <keyword>_field / <keyword>_param.

2. operationId collisions blocked whole documents — T6's strict-error policy was correct per spec but real-world docs (arcade, cal-com, telnyx, val-town) often violate it. Auto-disambiguate by suffixing with HTTP method, with a stderr warning.

3. Extensions rejected non-x-* keys — produces, in, type, density, title, description were observed on objects where they don't belong. Loosened Extensions to accept any leftover key and exposed non_extension_keys() so silent drops remain visible.

4. exclusiveMinimum: bool vs number — 3.0/Swagger used bool; 3.1 uses number. Modeled as a bool | f64 enum.

5. Vec<serde_json::Value> Ident panic — generate_array_item_type split on :: and produced strings with angle brackets that aren't valid Rust idents. Now parses via syn::parse_str::<syn::Type> first.

6. enum variant collisions on signed numbers — 1 and -1 both produced Variant1. Prefix negatives with Neg (e.g. VariantNeg1).

7. Twilio-style filter param ident collisions — StartTime, StartTime<, StartTime> all snake-cased to start_time. Map <, >, <=, >= to _lt/_gt/_lte/_gte in sanitize_param_name.

8. Version gate bypassed by the TOML config flow — generate subcommand has its own pipeline that bypassed cli::run_generation_cli. Mirrored the check so Swagger 2.0 specs error early with a clear hint instead of failing later inside the deserializer.

CI: spec-compile gold list

The spec-compile job now verifies a gold list of 20 specs that compile cleanly: anthropic, asana, browserbase, cartesia, cerebras, coda, coingecko, digitalocean, groq, imagekit, launchdarkly, meta-llama, openai, resend, runway, spotify, terminal-shop, twilio, val-town, writer.

Local scripts/spec-compile.sh (no args) still runs the full corpus for development. The script auto-discovers specs/*.{yaml,json}, skips Swagger 2.0 (gitea), and supports SPEC_COMPILE_PARSE_ONLY=1 for quick generator-only checks.

What's still failing (34 specs, tracked in #14)

Error class	Affected	Root cause
E0308 mismatched types	stripe (1632), box (99), vercel (111), …	Codegen emits wrong types in some property paths
E0428 name defined multiple times	github, gcore	Type/variant deduplication misses some collisions
E0117 orphan rule	stripe, gcore	Tries to impl foreign traits on foreign types
E0072 recursive type infinite size	snyk	Missing Box<T> for some recursive references
E0382 use of moved value	(now reduced)	String reused in generated method body

Each is a focused follow-up. Adding a fix → spec moves from CHECK-FAIL → PASS → can be added to the gold list.

Test plan

• cargo test --tests — 205/205 pass
• cargo clippy --all-features -- -D warnings — clean
• cargo fmt --check — clean
• scripts/spec-compile.sh against all 54 specs — 20 compile, 34 fail with categorized errors
• CI spec-compile job gates on the 20 working specs

🤖 Generated with Claude Code