fix(LTX25-DEVICE-RESIDENCY): name the load prologue load.open, because a floor wide enough to hide it is a mute switch (#1439)

[localai-bot]

CHECK(leaves >= 0.95 * wall) in ltx2 video: a render through the ABI emits a phase table that SUMS to wall reds main intermittently, and it failed at
0.702349 >= 0.702944 — 0.08% short. Two earlier sessions deliberately left it
alone and recorded why: 91% of its residue is ONE contiguous ~11.3 ms interval at
the head of the load span, before load.dit opens. It is a real startup phase.
Any slack big enough to stop the flapping would exceed 12 ms and would hide it.

So this names the phase instead of widening the floor. The assertion is
byte-for-byte unchanged. ltx2_video.cpp opens phase::Scope open_phase("load.open") immediately after the load span and closes it
immediately before the load.dit block, over the prologue the probe already
identified: device resolution, the platform probe, the two refusals, the
device-byte probe install, the DiT safetensors header open and the LoRA option
parsing. Scope::Close rather than a nested block is the shape
render_phase_log.h documents for this driver, whose regions are sequential
statements in one function.

Red first, and the red IS the reachability mutation. The test change was built
alone, so the red arm is this tree with the gate asking for the name and no
production scope to answer it: 447 assertions, 1 failed, the phase table names no 'load.open' phase. With the scope: 459 assertions, 0 failed. Deleting the
production call site is precisely that first arm.

A second mutation pins the CLOSE and not only the OPEN. Sinking
open_phase.Close() below the load.dit block compiles and keeps the SUM case
green at 459/459, and is caught instead by the #1440 nested-record rule: the
carrying-phases case reds at 595/1 with 'load.dit' is emitted NESTED. Tree
md5-verified restored between arms.

The residue, six runs per arm on one binary each, Release CPU: median
unaccounted_seconds 2.91 ms before, 0.94 ms after; load.open is emitted as a
leaf at t=0.000 and measures about 1.3 ms.

Two limits on that number, in the spec and repeated here because they bound the
claim. The box was at loadavg 24-31, so the render's wall is 7-10 s rather than
the 0.22-0.58 s of #1439's population — the flattering regime the issue itself
identified, where a slower render passes. And the 11.3 ms head is not reproduced
at this build type with the fixture warm; the prologue is about 1.3 ms here. The
direction and the mechanism are measured; the magnitude at which the floor bites
is inherited from the earlier probe, and a quiet-box run at fixture scale stays
owed.

The order case gains load.open as its first entry, so a swapped scope name reds
there the way load.dit and load.prompt_embeds already do.

Not owed: a docs/models/ltx-2-5.md edit. The spec said this change owes the
phase names published there. That document publishes none — ## Inspect a render
names sum_leaf_seconds, unaccounted_seconds and two switches and no
individual phase. Checked rather than assumed.

Refs #1439, which stays OPEN. The repair is measured; the regime where the
floor bites is not, and one quiet-box run at fixture scale is what closes it.
That run is owed by LTX25-DEVICE-RESIDENCY and is named in its ## Owed row.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5 [Claude Code]