Experiment #2 by semantic owner: the rotary schedule, and two declarations nothing reads

[chrishayuk]

Fixes 2, 3 and 4 of census experiment #2, by semantic owner. Three commits, three different treatments — and in every case the recorded classification was wrong.

---

1. no_rope_layers — the key means the opposite of its name

SmolLM3 and Llama 4 both declare which layers encode position with no_rope_layers plus a no_rope_layer_interval fallback. SmolLM3's config documents it as:

A 1 at an index position indicates that the corresponding layer will use RoPE, while a 0 indicates that it's a NoPE layer.

and both references read it as self.use_rope = config.no_rope_layers[layer_idx].

A reader who trusts the name inverts the entire schedule. On SmolLM3-3B that is 27 of 36 layers rotated that must not be, and 9 left unrotated that must be — and the model still emits fluent text, because NoPE layers are a minority and attention degrades gracefully. Nothing but a test pinned against the reference catches it.

Honoured exactly once, in PositionPolicy::rope_enabled_by_flag, beside the layer_rope_theta zero sentinel it is the sibling of.

Not an alias, as recorded: ALIAS_KEYS requires the canonical spelling present in the same config, and SmolLM3 declares no layer_rope_theta at all. The capability was shared; the spelling had no reader.

Placed in the trait default, by the criterion rotary_policy's doc already states — two unrelated families declare it with identical semantics, so it is a config fact rather than a family fact. The schedule is asked before the rotary shape and composes with it, so a scheduled layer still picks up YaRN, a partial rotary, or a per-layer theta.

no_rope_layer_interval is a fallback generator, not a restatement: both references consult it only if no_rope_layers is None. Carried at Parsed, because on a checkpoint declaring both it contributes nothing to the graph.

Two malformed cases decline to answer, both in the direction that preserves existing behaviour: a short mask says nothing about the layers past its end, and an interval of zero is not a schedule. Clippy's manual_is_multiple_of is what surfaced that my first zero-guard fell the wrong way — it would have turned a malformed declaration into a silently NoPE model.

Calibrated: inverting the flag fails 5 of 7 gates including the real 36-layer schedule. The two survivors declare no schedule at all — the control.

---

2. rope_interleaved and use_mrope — vestigial, and read anyway

A gap kind distinct from the declared-disabled one. use_mrope appears nowhere in transformers 5.5.0; rope_interleaved (exact spelling) appears nowhere either — SmolLM2-135M declares it under model_type: llama, which has no such field. Unlike use_sliding_window, which HF genuinely consults, there is no upstream behaviour to agree with.

That makes reading them more important, not less. Neither names a cosmetic fact:

rope_interleaved   which dimensions rotate against which partner
use_mrope          whether position is one axis or three

Each is checked against the effective policy, never against itself:

• rope_interleaved against ROPE_PAIRING_INTERLEAVED, a new constant naming this build's one pairing — split-half, (x[i], x[i + half]).
• use_mrope against whether mrope_section + mrope_interleaved actually resolved an MRope policy. A probe echoing the flag back would agree with a checkpoint claiming multi-axis while carrying no geometry to build one from.

the_executor_pairs_split_half keeps the constant honest in both directions: the expected partner is derived from the constant, so flipping it flips what the test demands and it then fails against an unchanged executor. Calibrated by making the executor interleaved for real — the gate goes red.

The parser reads these and stops. Reading a flag must not become acting on it: this build has one pairing and the planner reports the disagreement, where a parser that quietly switched operators would turn a reportable mismatch into the silent wrong answer the report exists to prevent.

---

3. Probe coverage

Both new carriage probes from the previous PR were declared and unexercised, which took carriage.rs under the 90% floor — and left the claim that these rules are checkable resting on nothing. Each now gets paired arms. The null arm is the one worth having: a stack that rotates nowhere answers null for position_embedding_type, and a checkpoint declaring rope over it must stop reading as satisfied — the granitemoehybrid case, reproduced on a fixture.

---

Census

HuggingFaceTB/SmolLM3-3B      3 → 1 blocking
HuggingFaceTB/SmolLM2-135M    3 → 2 blocking
Qwen/Qwen2.5-0.5B             2 → 1 blocking
everything else               unchanged

All three residuals are pre-existing and outside this cluster: pretraining_tp and is_llama_config are training/identity facts, and Qwen2.5 keeps the honest sliding_window mismatch — declared 32768, carried null, because the graph records the effect and not the reason.

Gates

fmt --all --check · check --workspace --all-targets (four new ModelConfig fields) · clippy with each touched crate's own CI flags, including larql-compute with and without --all-features · full suites for larql-models, larql-compute, larql-vindex, larql-server · coverage policy for all three gated crates, re-measured after llvm-cov clean.