fix(desktop): emit camelCase config-write payload fields

[Chessing234]

Fixes #6015.

ConfigWriteMechanism is internally tagged and carried only rename_all = "camelCase". On an internally tagged enum that renames the variants, never the variants' fields, so the payload went out snake_case:

{"type":"respawnWithEnvVar","env_key":"K"}
{"type":"acpSetConfigOption","config_id":"c"}
{"type":"acpSetSessionModel"}
{"type":"gooseNativeConfigWrite","config_key":"g"}
{"type":"readOnly"}

against envKey / configId / configKey in desktop/src/shared/api/types.ts:615-620. That output is a probe run of the real module before the fix, not a reading of the code.

What makes it read as correct is the asymmetry: the variant names rename fine, so the type discriminant and every switch (writeVia.type) behave; and the enclosing NormalizedField's own fields (writeVia, overriddenValue, isRequired) rename fine too, because rename_all does apply to struct fields. Only the variant's field is wrong.

Adding rename_all_fields = "camelCase" fixes it. rename_all_fields appeared zero times in desktop/src-tauri before this.

Severity, stated plainly: latent, not currently user-visible. Nothing in desktop/src/** reads .envKey/.configId/.configKey off a writeVia — AgentConfigPanel.tsx is the only RuntimeConfigSurface consumer and never touches the field, and no Rust code deserializes the type either. The write-back path these fields exist for is not wired yet. The hazard is for whoever wires it: invokeTauri<T> is an unchecked cast, so they get undefined with a green tsc.

I also carried the attribute onto ConfigFieldType. Its only payload field is options, single-word, so that half is not a fix — it is the attribute the next multi-word field would silently need.

One divergence from the issue's suggested step 3: the 20 e2eBridge.ts sites already emit camelCase, and camelCase is the contract, so they are correct as written — changing them would have been wrong. What they lacked was provenance, since agreeing with api/types.ts while the backend emitted something else is exactly what let this sit. They now name the Rust test that pins the bytes.

Tests (wire_format_tests, 4 cases, whole-value not key-set — a key-set assertion still passes when a variant name regresses):

• every variant against the TypeScript spelling;
• the nested NormalizedField, which is the shape the renderer actually receives;
• a camelCase round-trip plus an assertion that the old env_key spelling is now rejected, so a revert cannot quietly keep deserializing;
• ConfigFieldType::Enum.

Removing rename_all_fields again turns three of the four red.

Verified locally: full Tauri library suite 2444 passed / 15 ignored / 0 failed; cargo clippy --manifest-path desktop/src-tauri/Cargo.toml --workspace --all-targets -- -D warnings clean; cargo fmt --manifest-path desktop/src-tauri/Cargo.toml --all -- --check; in desktop/: pnpm typecheck, pnpm check, pnpm test 4954 passed; git diff --check. Not run: the app itself — there is no UI path to this field yet, which is the same reason the bug is latent.

[themiguelamador]

I found and fixed two minor review issues:

• P3 — Keep the patch scoped to the broken wire contract. rename_all_fields was also added to ConfigFieldType, but its only payload field (options) is already a single camel-case word. That change has no effect today and silently chooses serialization behavior for future fields unrelated to this bug, so I removed it and its non-regression test.
• P3 — Assert JSON semantics, not serializer property order. The new tests compared exact serialized strings, coupling the contract tests to object-key order even though JSON object order is not semantic. I changed them to whole-value serde_json::Value comparisons, which still verify every key and value (including nested fields) without overconstraining the encoder.

Fix branch: https://github.com/Complear/buzz/tree/review/pr-6062-fix
Commit: d0f57d1de

Validation: full Tauri library suite (2,443 passed, 15 ignored), strict Tauri workspace/all-target clippy, focused wire-format tests, cargo fmt, desktop typecheck, full desktop checks, and git diff --check all passed. The desktop checks reported only the four pre-existing advisory Biome diagnostics outside this PR.

[Chessing234]

Both applied, one commit each (3ae0add, 935d6a4). (Complear/buzz 404s, so written from your description.)

P3 — scope. Verified inert before removing it: ConfigFieldType's other three variants are unit variants, and api/types.ts declares the enum as { type: "enum"; options: string[] }, so rename_all_fields changes no byte the renderer sees today. Removed it and its test — that test was explicitly a shape pin, not a non-regression, and nothing it asserted could have failed. ConfigWriteMechanism keeps the attribute, where it is the fix.

Worth stating the trade-off you're accepting, since it's the same footgun this PR exists to close: a future multi-word field on ConfigFieldType will serialize snake_case and invokeTauri<T>'s unchecked cast won't catch it. I agree that's not this patch's job to pre-empt.

P3 — JSON semantics. Agreed, key order isn't semantic. The two serialization tests now compare whole serde_json::Values built with json!, which still catches a renamed variant or field. The round-trip test keeps its string literals deliberately: there the exact input bytes are the point, since it asserts the renderer's shape is accepted and the pre-fix snake_case spelling isn't.

Verification: full Tauri lib suite (2,443 passed, 15 ignored), strict workspace/all-target clippy, and cargo fmt --check.