fix(requestflag): merge inner flags into untyped array literals

[feiiiiii5]

Summary

For a body parameter that is a nullable array of objects, codegen emits an untyped
requestflag.Flag[any] plus InnerFlag children marked OuterIsArrayOfObjects. Combining
an array literal on the outer flag with any inner flag silently discards the inner value:

openai fine-tuning:jobs create \
  --model gpt-4o-mini --training-file file-abc123 \
  --integration '[{type: wandb}]' \
  --integration.wandb '{project: my-project, entity: me}'

request body built from ExtractRequestContents:

{"model":"gpt-4o-mini","training_file":"file-abc123","integrations":[{"type":"wandb"}]}

wandb is gone — no error, no warning, and --integration.wandb still counts as supplied,
so nothing downstream notices. Expected {"integrations":[{"type":"wandb","wandb":{…}}]}.

Problem

cliValue.SetInnerField dispatches the slice case on the static element kind
(internal/requestflag/requestflag.go):

case reflect.Slice:
	if flagValReflect.Type().Elem().Kind() != reflect.Map {
		return
	}
	...
	lastElement := flagValReflect.Index(sliceLen - 1).Interface().(map[string]any)

A Flag[any] outer holds a []any as soon as it is set from a JSON or YAML array literal,
and []any's element kind is reflect.Interface, not Map. Every inner-field assignment
therefore returns at that guard, while Flag.SetInnerField goes on to set
hasBeenSet = true. The typed outer Flag[[]map[string]any] (element kind Map) takes the
merge path — which is the behavior the existing TestInnerFlagAfterNullArrayElement cases
pin, including "merge into the trailing element" and "repeated field starts another element".

Two details in the surrounding code say the []any shape was meant to work: the append
switch already has a case []any: arm that the guard made unreachable, and
innerFieldIsSet in stdinprovenance.go already reads the trailing element of a []any
with a checked assertion.

Affected parameters, all five Flag[any] with array-of-objects inner flags (verified in
--help of the built binary): fine-tuning:jobs create --integration,
responses create --context-management, beta:responses create --context-management,
beta:threads:runs create --additional-message, beta:threads:messages create --attachment.

Live request capture

Same command against a local recorder on 127.0.0.1:4011 (synthetic data, no live endpoint),
binary built from 0169bff vs from this branch:

$ openai --api-key test --base-url http://127.0.0.1:4011 fine-tuning:jobs create \
    --model gpt-4o-mini --training-file file-abc123 \
    --integration '[{type: wandb}]' --integration.wandb '{project: my-project, entity: me}'

build	POST /fine_tuning/jobs body, integrations
0169bff	[{"type": "wandb"}]
this branch	[{"type": "wandb", "wandb": {"entity": "me", "project": "my-project"}}]

Both runs exit 0 and get a 200 — the only signal that the field was lost is the body itself.

Fix

Dispatch elements by their dynamic type, and merge only into a trailing object:

switch flagValReflect.Type().Elem().Kind() {
case reflect.Map, reflect.Interface:
default:
	return
}

sliceLen := flagValReflect.Len()
if sliceLen > 0 {
	lastElement, isObject := flagValReflect.Index(sliceLen - 1).Interface().(map[string]any)
	if isObject {
		… unchanged …
	}
}

• Trailing map[string]any → existing rule (merge unless it already carries the field;
  fill a nil map in place).
• Anything that is not an object — a scalar element, an untyped null element, or a trailing
  element that already has the field → append a new object element, i.e. the same
  "repeated field starts another element" rule the typed path already uses. Nothing the user
  typed is discarded either way.
• The unchecked Interface().(map[string]any) becomes checked, so widening the guard cannot
  turn the assertion below it into a panic path.

No new limits, no error paths, no signature changes: SetInnerField's contract and the
Flag/InnerFlag wiring are untouched. This is the remaining half of the gap #162 closed:
that change made a null element inside a typed []map[string]any work, while the guard
above still drops every inner flag whose outer is an untyped Flag[any].

Tests

New internal/requestflag/innerflag_untyped_array_test.go, 8 cases. Seven run through the
real cli.Command.Run → ExtractRequestContents path (the same map the request body is
built from), one drives Flag/InnerFlag directly to mirror the existing null-element
test:

vector	expected body
--entry '[{"name":"earlier"}]' --entry.description details	[{"name":"earlier","description":"details"}]
--entry '[{"name":"earlier"}]' --entry.name demo	[{"name":"earlier"},{"name":"demo"}]
--entry '[{}]' --entry.name demo --entry.description details	[{"name":"demo","description":"details"}]
--entry '- {name: earlier}' --entry.description details (YAML flow)	[{"name":"earlier","description":"details"}]
--entry '["plain"]' --entry.name demo	["plain",{"name":"demo"}]
--entry '[null]' --entry.name demo	[null,{"name":"demo"}]
--entry '[]' --entry.name demo	[{"name":"demo"}]

On base 0169bff (same test file, requestflag.go restored from git show HEAD:…) all 8
fail, each dropping the inner field:

--- FAIL: TestInnerFlagAfterUntypedArrayLiteral/merge_into_the_trailing_element
	expected: …{"description":"details", "name":"earlier"}
	actual  : …{"name":"earlier"}

With the patch all 8 pass, and the pre-existing internal/requestflag suite passes
unchanged, so the typed []map[string]any / map[string]any vectors are unaffected.

For context on the syntax: inner flags are not listed in --help, but they are codegen'd
for exactly these parameters and are the style the generated inner flags subtests drive
(TestFineTuningJobsCreate in pkg/cmd/finetuningjob_test.go passes --integration.type
and --integration.wandb). Those tests use only inner flags, which is why the regression
escaped: the lost field needs an array literal and an inner flag in the same command.

Validation

cwd .runtime/2026-09-20/oa/src, base 0169bff (= origin/main), go1.25.0 darwin/arm64,
network proxies unset, -count=1 where shown:

Command	Result
go test ./internal/requestflag/ -run 'TestInnerFlagAfterUntypedArrayLiteral' -v (unpatched)	8 FAIL
go test ./internal/requestflag/ -run 'TestInnerFlagAfterUntypedArrayLiteral' -v (patched)	8 PASS
go test ./internal/...	ok, 0 failures
go test ./internal/requestflag/ -race -count=1	ok, 1.416s
go test ./... -run '^$'	ok, all packages compile
go test ./pkg/cmd -count=1 with ./scripts/mock on 127.0.0.1:4010	62.9s, 1 failure: TestFilesCreateCLICancelClosesStalledFIFO
go test ./pkg/cmd -run TestFilesCreateCLICancelClosesStalledFIFO on base, patch reverted	same failure → pre-existing, unrelated (FIFO cancellation, multipartresource_unix_test.go)
go vet ./...	no output
./scripts/lint	rc=0
gofmt -l internal/requestflag/	no output
go mod verify	all modules verified
go build -o /tmp/oai ./cmd/openai (patched and git show HEAD: variants), each run against a local 127.0.0.1:4011 recorder	bodies as tabulated under "Live request capture"
GOOS=windows GOARCH=amd64 go build ./..., GOOS=linux GOARCH=arm64 go test -c ./internal/requestflag/	ok

The mock-server dependency was reviewed before running it: scripts/steady/manifest.json
pins the Steady commit and Deno runtime, ./scripts/steady/install verified the source and
runtime digests, and the server bound to 127.0.0.1:4010 with synthetic data only.
./scripts/test itself was not run (it additionally cross-compiles the whole test tree); the
pkg/cmd suite above is the mock-backed part that covers this change.

internal/requestflag/requestflag.go is handwritten behavior per AGENTS.md, and the new
test file is wholly handwritten, so no generated-owned file changes and nothing in the
Castiron generation baseline moves. The change is in request parsing, so CODEOWNER review is
expected.

AI-assisted: an AI coding agent produced this change under the account owner's standing
instruction for this repo, ran every command above and read its output. No human reviewed
the diff before it was opened.