test(lfm2): pin the warm-continue comparison to answer-region decisions - #152

The problem, corrected

main failed e2e lfm2 on three consecutive runs (#148, #149, #150), and it is not a random flake — it is host-dependent determinism:

[warm-cont] turn1:             num_tokens=2 finish=length raw_text="</think> Ok"
[warm-cont] FLAT turn1:        num_tokens=2 finish=length raw_text="</think> Ok"     ← arms identical before the boundary
[warm-cont] turn2 WARM:        num_tokens=32 cached_tokens=23 finish=length
[warm-cont] turn2 FLAT oracle: num_tokens=32 cached_tokens=23 finish=length          ← same reuse, same length
WARM raw_text="<think> Okay, let's see. The user said to keep going from where I left off, but …"
FLAT raw_text="<think> Okay, let's see what the user is asking here. They said, \"Now …"

first_diff_byte=Some(23) is identical on all three red runs, and the same commit passes locally. So the carry path is fine, the arms are comparable, and what split is the first free-form decision inside the thinking span — the near-tie that this file already documents as flippable by the accepted ~1-ULP paged-vs-flat kernel-stack class (models/lfm2/attention.rs, and the "Deliberately NOT asserted: byte-equality of the answer texts" note on the memory-probe test). Byte parity was being asserted over a coin flip.

I first tried downgrading that assertion to a diagnostic; an adversarial review rightly rejected it — "a stale or incorrect conv tensor can preserve turn-1 output, reuse 23 tokens, and generate 32 incorrect tokens", so invariants cannot replace the bite. This is the reviewer's own recommended alternative instead.

The change (one file)

• Turn 2 is now budget-forced, like turn 1. Both arms therefore commit the same forced </think> first and the compared tokens are answer-region ones, not a sentence-opening word choice. The boundary under test (turn-1/turn-2 carry, ContinuedLivePrefix, 23-token prefix) is untouched.
• Byte parity stays fatal — the assertion is unchanged in strength.
• Three invariants added, which the fixture was missing: the arms' turn-1 outputs must be equal (turn 1 is cold, so a mismatch is a fixture lottery, and the message says so instead of sending a reader into the continue path), the reuse lengths must be equal, and the committed token counts must be equal (the qwen3 sibling already asserts that one).

Verification

What is still not proven

The compared region is now answer-region, but a near-tie there is still possible in principle, so this is a fixture pin validated on the affected host, not a proof that the split was harmless rounding. The durable instrument is the one models/qwen3/model.rs uses for its mixed-step gate: logits at turn-2's first forward within test_support::bf16_scaled_tolerance, plus top - second > 2 × tolerance. That needs a logits readout a test can reach — ChatResult carries none and Lfm2Inner is crate-private — so it is a follow-up, not something to fake with a looser text comparison here.

This clears the model-test gate that has been skipping Publish on main.