feat(inference): parse Qwen3.5 vision config fields (ADR-069 S1)

[ohdearquant]

Authored by Claude (Anthropic agent) on behalf of @ohdearquant.

Summary

ADR-069 stage S1: parse the vision-language checkpoint fields that
qwen35_config.rs was previously dropping, without wiring them into the
forward pass.

• New VisionConfig struct (ViT depth, hidden_size, num_heads,
  patch_size, spatial_merge_size, out_hidden_size,
  temporal_patch_size, num_position_embeddings, in_channels,
  deepstack_visual_indexes), parsed from the top-level vision_config
  object in config.json. None on Qwen35Config for text-only
  checkpoints.
• Four vision token ids (image_token_id, video_token_id,
  vision_start_token_id, vision_end_token_id), threaded from the
  top-level config.json the same way tie_word_embeddings already is.
• RopeParams gains mrope_section: Option<Vec<usize>> and
  mrope_interleaved: Option<bool>, parsed from
  text_config.rope_parameters alongside the existing rope_theta and
  partial_rotary_factor handling (unchanged).

All five new Qwen35Config fields are additive and default to None;
every preset constructor and the existing config.json parse path are
otherwise unaffected. Consumption by the forward pass (ViT encoder,
merger, decoder M-RoPE) lands in later ADR-069 stages -- these fields are
parsed and stored only.

Note on scope

Qwen35Config is a 31(+5)-field struct built via fully-exhaustive struct
literals (no ..Default::default() spread) in several test helper
functions across the crate. Adding new fields to the struct is a source
change that, by Rust's own construction-exhaustiveness rules, requires a
matching field at every such literal. This PR adds the 5 new fields (all
None) at each of those existing test-helper literals -- purely
mechanical, no logic changes -- so the crate keeps compiling under
default, metal-gpu, and all-features builds. The real feature work is
entirely contained in qwen35_config.rs.

Verification

cargo test -p lattice-inference qwen35_config::

test model::qwen35_config::tests::test_qwen35_0_8b_config_json_fixture_parses ... ok
test model::qwen35_config::tests::test_text_only_config_has_no_vision_fields ... ok
test result: ok. 52 passed; 0 failed; 0 ignored; 0 measured; 1456 filtered out; finished in 0.00s

cargo clippy -p lattice-inference -- -D warnings

    Finished `dev` profile [unoptimized + debuginfo] target(s) in 2.03s

(also clean under --all-targets, --features metal-gpu, and
--all-features)

cargo test -p lattice-inference --lib

test result: ok. 1500 passed; 0 failed; 8 ignored; 0 measured; 0 filtered out; finished in 44.67s

cargo fmt -p lattice-inference -- --check is clean.

The round-trip test parses the real released 0.8B config.json fixture
(already committed at crates/inference/tests/fixtures/qwen35_0_8b_config.json)
and asserts every vision dimension, all four token ids, and
mrope_section == [11, 11, 10] / mrope_interleaved == true. A second
test asserts a text-only config.json (no vision_config, no
rope_parameters) still parses cleanly with every new field resolving to
None, proving the plain-text decoder path is unaffected.

Follow-up

Consuming these fields (ViT forward pass, patch merger, image-token
injection, decoder M-RoPE) is out of scope here and lands in later
ADR-069 stages.

[github-actions]

E2E Parity Report

PASS: all 4 prompts match within their respective match windows

Prompt	Window	Agreement	First Diff	HF tok/s	Lattice tok/s	Verdict
The capital of France is	3	15/15	none	0.5	2.8	PASS
In the year 2024, artificial intelligence	3	15/15	none	0.4	2.4	PASS
`def fibonacci(n):

if n <= 1:
    return n
return` | 3 | 15/15 | none | 0.4 | 2.0 | PASS |

| def merge_sort(arr): """ Merge sort implementation. | 2 | 15/15 | none | 0.4 | 0.2 | PASS |

The capital of France is

• HF: Paris.
  The capital of France is Paris.
  The capital of France
• Lattice: Paris.
  The capital of France is Paris.
  The capital of France

In the year 2024, artificial intelligence

• HF: (AI) has become a significant part of the global economy. It is
• Lattice: (AI) has become a significant part of the global economy. It is

def fibonacci(n): if n <= 1: return n return

• HF: fibonacci(n-1) + fibonacci(n-2)

print(fib

• Lattice: fibonacci(n-1) + fibonacci(n-2)

print(fib

def merge_sort(arr): """ Merge sort implementation.

• HF: merge_sort(arr):
  """
  Merge sort implementation.

• Lattice: merge_sort(arr):
  """
  Merge sort implementation.

[github-actions]

E2E Parity Report

PASS: 3/4 gating prompts match; 1 known divergence (#535) excluded from the verdict

Prompt	Window	Agreement	First Diff	HF tok/s	Lattice tok/s	Verdict
The capital of France is	3	4/4	none	0.4	3.9	PASS
In the year 2024, artificial intelligence	3	4/4	none	0.2	7.1	PASS
`def fibonacci(n):

if n <= 1:
    return n
return` | 3 | 4/4 | none | 0.1 | 5.9 | PASS |

| def merge_sort(arr): """ Merge sort implementation. | 2 | 0/4 | pos 0 | 0.1 | 0.8 | KNOWN-DIVERGENT (#535) |

The capital of France is

• HF: Paris.
  The
• Lattice: Paris.
  The

In the year 2024, artificial intelligence

• HF: (AI) has
• Lattice: (AI) has

def fibonacci(n): if n <= 1: return n return

• HF: fibonacci(n-1
• Lattice: fibonacci(n-1

def merge_sort(arr): """ Merge sort implementation.

• HF: merge_sort(arr):
• Lattice: main():