fix(features): the last two struct-drift selections — depth_slice and stream_options

[noahgift]

What

The last two mechanical items from the 430-pair (crate, feature) sweep (#3262). Both are struct drift: a field was added upstream and the one initializer in the tree was never updated, because no gate ever builds these selections.

crate	field	why None is the right value
aprender-present-lib gpu	wgpu::RenderPassColorAttachment::depth_slice	It selects one slice of a 3D texture. The view here is a 2D surface texture, for which None is the only valid value. This is the only RenderPassColorAttachment in the workspace, so there was no sibling to copy.
aprender-test-cli llm	jugar_probar::llm::ChatRequest::stream_options	Its own doc settles it: the field "must be absent when stream is not true" because OpenAI rejects it outright on a non-streaming request — and this call site sets stream: Some(false). So None, not a copy of what wire_request sends on the streaming path.

With #3263 and this PR, every red pair in the matrix is fixed except the nine owned elsewhere: seven aprender-distribute selections (#3257 — its tokio gate is right, but executor/cpu.rs also uses num_cpus, which only the cpu feature pulls; verified against that PR's head) and two cuda selections (#3179 — a driver-API decision).

Complexity debt paid down

aprender-test-cli's one-line fix was blocked by the pre-commit complexity gate on six PRE-EXISTING violations in the same file, none in code this PR wrote. Paid down here, as the gate requires — all behaviour-preserving:

execute_llm_load        cyc 36 / cog 61  ->  under
execute_llm_score       cyc 40 / cog 91  ->  under
execute_llm_sweep       cog 40           ->  under
execute_llm_experiment  cog 36           ->  under

The largest single win is render_card: every --by-* section of execute_llm_score repeated the same three-arm match args.format verbatim and only the two formatter functions differed, so they are passed in — nine sites collapsed. The rest are block extractions: print_load_summary (~140 lines that decide nothing — they read result and print), print_tail_analysis, print_quality_validation, health_gate (GH-37's four outcomes, each still naming the endpoint and the bypass flag), detect_saturation (no baseline ⇒ never saturated; the reason string is byte-identical), level_status (saturation still wins over zero-quality), pareto_frontier, and experiment_status.

Verified

• both selections cargo check clean, and both crates' defaults clean
• aprender-test-cli --lib 753/753, aprender-present-lib --lib 60/60 — and all 10 of llm.rs's #[test] are still present, checked explicitly rather than inferred from the total
• clippy finds nothing in llm.rs; the five unwrap() warnings in this crate are pre-existing and in files this PR never touches (comply, load_testing, statistics, stress)
• no unwrap() added: git diff | grep '^+.*\.unwrap()' is 0

no-close: #3262 is the umbrella measurement and stays open until the last nine pairs are green; #3257 and #3179 are owned elsewhere.

🤖 Generated with Claude Code

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=3265 head=cf32b8da85bee403ee7d0e2d78bb1c3141cd9e41 verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.