P7 follow-up: GptService 6-arg ctor binary compat + ImageQuality "high" telemetry hardcoding

[cyberprophet]

Background

Follow-ups from PR #97 review (Codex narrowing, both P2 — non-blocking).

1. 6-arg GptService(ILogger, string, OpenAIClientOptions, string?, string?, string) constructor binary compatibility

Before 0.16.4, the custom-endpoint constructor had this signature (6 source params, 3 with defaults). 0.16.4 widened it to 7 params by adding an optional imageGenerationModel = null at the end. C# treats optional params as the same source signature, but emits a different IL token — assemblies compiled against ≤0.16.2 that explicitly call the 6-arg overload would hit MissingMethodException at runtime when loaded with 0.16.4-0.16.6.

Current real-world impact: none. The only external NuGet consumer is P5 (creative-server), and P5 always recompiles against the latest pinned reference, so the 6-arg IL token is never sought. Filed as belt-and-braces in case a future external consumer adopts ShareInvest.Agency.

Fix sketch: add an explicit 6-arg overload that delegates to the 7-arg with imageGenerationModel: null. May require removing the default on imageGenerationModel in the 7-arg overload to disambiguate; verify P5 call sites first (only Server/Extensions.Services.cs:155 and a couple of test files use new GptService(...) and they take the 2-arg form, so widening is safe).

Reference: PR #97 review by Codex on commit fda8414d87 — comment ID 3135 series.

2. ApiUsageEvent telemetry hardcodes ImageQuality: \"high\" for StudioMint

Even though 0.16.6 removed ImageEditOptions.Quality (so the edit endpoint now uses its default auto tier), GenerateSingleShotAsync still emits new ApiUsageEvent(... ImageQuality: \"high\" ...) for every successful shot. Telemetry / billing analytics are wrong by exactly one tier for the affected period.

Fix sketch: drop the \"high\" hardcode and pass the actual configured quality (or null while we're on the SDK workaround). When the OpenAI .NET SDK gets a fix for the quality serialization bug and we restore the option, this can flip back to a real value.

Reference: PR #97 review by Codex on commit fda8414d87 — comment on agency/OpenAI/GptService.StudioMint.cs:107.

Severity

Both P3 — no user-visible bug. Can ride the next non-urgent Agency bump (0.17.0).

[cyberchacha]

Triage Summary

• Type: enhancement
• Summary: Two P3 follow-ups from PR fix(studio-mint): drop unsupported output_format from edit options (0.16.3) #97 review: (1) missing binary-compat shim for the 6-arg custom-endpoint GptService constructor, and (2) ApiUsageEvent hardcoding ImageQuality: "high" when actual quality sent to OpenAI is auto (SDK workaround).
• Why this label: Both items are preventive improvements — no user-visible bug today. The author explicitly marks them P3 for the next non-urgent Agency bump (0.17.0).
• Missing info: none

Findings

Missing 6-arg binary-compat shim for custom-endpoint constructor

On origin/main, agency/OpenAI/GptService.cs:100 has a 7-param constructor:

public GptService(ILogger<GptService> logger, string apiKey, OpenAIClientOptions options,
    string? imageModel = null, string? exaApiKey = null, string providerName = "openai",
    string? imageGenerationModel = null)

Before 0.16.4 this was 6 params (no imageGenerationModel). C# emits a different IL method token for the widened signature, so any assembly compiled against ≤0.16.2 that calls the 6-arg form would hit MissingMethodException at runtime.

A binary-compat shim already exists for the simpler 4-param constructor (line 53), but no equivalent 6-arg shim exists for the custom-endpoint overload. Current real-world impact is nil (P5 always recompiles), but this is a correctness gap for any future external NuGet consumer.

ImageQuality: "high" hardcoded in telemetry despite quality not being set

On origin/main, agency/OpenAI/GptService.StudioMint.cs around line 120:

onUsage(new ApiUsageEvent(
    ProviderName,
    imageModel ?? "gpt-image-1",
    textInput,
    (int)(usage?.OutputTokenCount ?? 0),
    $"studio-mint:{shot.Id}",
    LatencyMs: (int)sw.ElapsedMilliseconds,
    ImageQuality: "high", ImageSize: "1024x1024",   // ← hardcoded
    ImageInputTokens: imageInput > 0 ? imageInput : null));

The ImageEditOptions block (line ~90) intentionally does not set Quality due to the OpenAI .NET SDK 2.10.0 serialization bug — the server receives no quality parameter and defaults to auto. But telemetry still reports "high", making billing/analytics overstate the tier for every StudioMint shot.

Suggested Approach

1. 6-arg shim: Add an explicit 6-param overload in GptService.cs that delegates to the 7-param constructor with imageGenerationModel: null. Remove the default value from imageGenerationModel in the 7-param overload to prevent C# overload resolution ambiguity. Verify P5 call sites (Server/Extensions.Services.cs:155 and test files) — they use the 2-arg form, so no impact.
2. Telemetry fix: Change ImageQuality: "high" to ImageQuality: null (or "auto") in GenerateSingleShotAsync to reflect the actual tier in effect while the SDK workaround is active. Add a // TODO: restore "high" when SDK quality serialization is fixed comment for traceability.
3. Test coverage: Existing ApiUsageEventTests should verify the new ImageQuality value. Add a unit test asserting the 6-arg constructor delegates correctly (e.g., imageGenerationModel is null, all other fields match).
4. Ship together in 0.17.0 as stated in the issue.

Risks or trade-offs: The 6-arg shim adds one more constructor overload to an already-growing constructor list — consider a builder/options-object pattern in a future cleanup pass. The telemetry change will cause a visible shift in analytics dashboards (from "high" to null/"auto") which downstream consumers should be aware of.
Estimated scope: small