Fix sampling option compatibility metadata for OpenAI reasoning models - #40

Summary

The provider previously advertised temperature, top_p, logprobs, and top_logprobs for every text-generation model. That made requirement-based resolution treat sampling-incompatible reasoning models as eligible, with failures deferred to the OpenAI API.

This PR makes the four sampling fields model-aware and keeps them consistent as a group. It also maps the AI Client SDK's native log-probability options into the Responses API, rejects explicit non-none reasoning combined with sampling before a request is sent, adds focused PHPUnit coverage, excludes test assets from both release-package paths, and adds read-only CI across PHP 7.4–8.4.

Compatibility matrix

“Advertise sampling” means the provider metadata advertises all four fields: temperature, top_p, logprobs, and top_logprobs.

Fine-tuned IDs are classified through their extracted base model while the original fine-tuned ID remains in ModelMetadata; tests cover an eligible ft:gpt-5.2:… and an ineligible ft:gpt-5.2-codex:….

The two versioned chat aliases are deliberately conservative exceptions to their documentation-derived chat classification. Live verification directly established only that temperature is rejected. Because provider metadata exposes the four sampling fields consistently, the grouped advertisement is withdrawn for those aliases; this does not claim that every sampling field was independently live-probed or rejected.

Future GPT-5 minor versions remain conservative until documented and tested. The reasoning-family matcher recognizes only the verified o1, o3, and o4 families; future o-series families are not auto-covered.

Request mapping and validation

• ModelConfig::getLogprobs() === true maps to include: ["message.output_text.logprobs"].
• A non-null ModelConfig::getTopLogprobs() maps to top_logprobs.
• logprobs: false leaves include unset.
• temperature, top_p, logprobs, and top_logprobs remain allowed with explicit reasoning.effort: none.
• Any of those fields combined with an explicit non-none reasoning effort raises InvalidArgumentException before the API request.

These names match the official Responses create reference, including message.output_text.logprobs and top_logprobs.

Sanitized live Responses probes

Date: 2026-07-18. Every request used input Reply with OK. and max_output_tokens: 16. No credential, header, request/response identifier, raw response, or account identifier is recorded.

Payload categories:

• temperature_default_none: temperature: 0.2, relying on the documented default effort.
• reasoning_none_top_p: reasoning.effort: none, top_p: 0.9.
• logprobs_top_logprobs: include: ["message.output_text.logprobs"], top_logprobs: 5.
• chat_temperature: temperature: 0.2 on the named chat alias.

temperature_rejected is a normalized report result, not an OpenAI error code. Availability/access/model-not-found responses do not become incompatibility rules. No reasoning_high_temperature probes were sent because none of the five reasoning-model baselines returned a 2xx accessibility confirmation.

Packaging and release scope

• Composer/git archives exclude tests/ and phpunit.xml.dist via .gitattributes.
• WordPress.org staging excludes the same assets via .distignore while retaining plugin.php and production source.
• PR-introduced symbols use @since 1.1.0; the plugin header remains 1.0.3 for the release process.
• The package keeps PHP >=7.4 and the existing AI Client SDK dependency line.

Verification

Local at a422636d52bc66696528bb1370c63c94eeafc934

• composer validate --strict --no-check-publish: pass.
• composer lint: pass; PHPCS 9/9 files, PHPStan 5/5 analyzed paths with no errors.
• composer test: pass; 35 tests, 178 assertions.
• git diff --check trunk...HEAD: pass.
• git status --short: empty.
• Composer archive: 15 entries, 0 tests/ or phpunit.xml.dist matches.
• WordPress.org staging: tests/ and phpunit.xml.dist absent; plugin.php and src/Metadata/OpenAiModelMetadataDirectory.php present.

GitHub Actions

Exact-head run: CI run 29640691457

The quality job runs composer lint and both release-package exclusion checks. All Actions use read-only repository permissions and immutable action SHAs.

Fixes #38