feat(graphile-llm): add inference usage logging to metering plugin

[pyramation]

Summary

Adds inline INSERT INTO usage_log_inference to the LlmMeteringPlugin, writing to the parent database alongside the existing billing record_usage calls. This is Approach 1 from the metering architecture discussion — simple, SQL straight up, no new packages.

What changed

config-cache.ts

• New InferenceLogConfig type (schema + table name)
• New INFERENCE_LOG_MODULE_SQL query to resolve the inference log module's schema/table from metaschema_modules_public.inference_log_module
• New resolveInferenceLogConfig() — checks if schema exists, queries module config, returns null if not provisioned
• getLlmBillingConfig() now resolves inference log config in parallel with billing + API key

metering.ts

• MeteringContext expanded with databaseId, actorId, inferenceLog
• MeteringOptions expanded with embeddingModel, chatModel, provider (for log entries)
• New InferenceLogEntry interface — typed payload matching usage_log_inference columns
• New logInferenceUsage() — parameterized INSERT, graceful skip if module not provisioned, catch logs warning but never throws
• meteredEmbed — logs after billing on both success and quota_exceeded paths
• meteredChat — logs after billing on both success and quota_exceeded paths

metering-plugin.ts

• buildMeteringContext now extracts actorId from jwt.claims.user_id and reads inferenceLog from the cached config entry
• Passes embeddingModel and chatModel to MeteringOptions

index.ts

• Exports logInferenceUsage, InferenceLogEntry, InferenceLogConfig

Design decisions

• Parent DB only — writes to the database the request is connected to via withPgClient. Child DB / platform dual-write is a TODO comment for future work.
• Graceful degradation — if inference_log_module is not provisioned, inferenceLog is null and logInferenceUsage returns immediately. If the INSERT fails, it logs a warning but never throws.
• Cached resolution — inference log config is resolved once per database_id (5-min TTL) alongside billing config, so there's no extra query per request.
• Quota exceeded logging — both embed and chat log status: 'quota_exceeded' events so the usage table captures denied requests too.

Depends on

• PR feat(graphile-llm): billing metering + agent discovery + REST streaming endpoint + env config #1192 (feat/llm-billing-metering) — base branch, provides metering.ts + config-cache.ts + metering-plugin.ts
• PR #1237 + #1239 (constructive-db, merged) — inference_log_module generator that creates the usage_log_inference table

Review & Testing Checklist for Human

• Verify the INSERT column list matches the inference_log_module generator output (16 columns: database_id through error_type)
• Verify logInferenceUsage never throws — the catch in both call sites + the try/catch inside the function itself should make this bulletproof
• Confirm that resolveInferenceLogConfig correctly guards against missing metaschema_modules_public schema
• Test with a database that has inference_log_module provisioned — verify rows appear in usage_log_inference after embedding/chat calls
• Test with a database without inference_log_module — verify no errors, metering/billing still works normally

Notes

• The provider field is not yet populated (always null) since the current LlmModulePlugin doesn't expose provider info on the build context. This can be wired in a follow-up.
• Future: graphile-usage-log plugin for generic multi-category logging with routing (platform/app/dual) — tracked in #901

Link to Devin session: https://app.devin.ai/sessions/2b5a29d83d3f478e8d3d972653b4879c
Requested by: @pyramation

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring