Many-grain density support (dense/streaky frames)

[marinerhemant]

Enables indexing of dense, streaky, many-grain frames — hundreds of overlapping grains
per image, no depth resolution — which the pipeline previously mishandled. This is the
core capability behind the July 2026 34-ID-E Ti-6Al-4V campaign (13 scans, ~150k frames).
Validated against multi-grain synthetic ground truth.

1. Robust blur sizing (laue_index/preprocess.py)

The auto matching-blur sigma was sized from the absolute minimum neighbour distance,
which collapses on dense frames: fragments of a single streak sit a few px apart, shrinking
sigma to ~1 px and leaving the 0.4° orientation grid unable to reach spots up to ~18 px from
a grid seed. Recovery dropped to ~15% on a 150-grain synthetic frame. Fragments of one
streak should merge — they are one grain's reflection — so size from the 10th percentile
of neighbour distances. This changes the default preprocessed output; golden fixtures updated
to match.

2. GaussSigmaMax parameter (laue_stream_utils.py + preprocess.py)

Optional cap on the auto sigma. On dense frames a large blur lights so much of the detector
that chance matches flood the coarse gate and bury true grains; iterative-peel drivers start
tight and widen per pass. Default 0.0 = no cap, so existing behaviour is unchanged unless
the param is set.

3. MAX_MATCHES 100k → 4M (LaueMatchingGPU.cu, LaueMatchingGPUStream.cu)

The kernel appends matches in arrival order, so overflow discards candidates arbitrarily —
true grains included — before the top-score merge can rank them. 100k overflowed on
many-grain frames. 4M × 12 B × MAX_STREAMS is < 200 MB on device and pinned host.

4. Torn-row tolerance (read_solutions)

Tolerates truncated rows (daemon terminated mid-write on a dense frame), dropping malformed
rows with a warning rather than failing the whole parse.

Plus

GenerateSimulation.py gains multi-grain synthetic frames (orientations_to_recips) — the
known-orientation, known-count ground truth used to measure recovery above.

Tests

Char suite and the streaming regression tests (#5) pass. The .cu change needs the GPU
build (build-gpu) to verify — the CPU build does not exercise it.

Deliberately not included: the laue_torch README bullet and its torch/midas-hkls
requirements, held until the associated publications go through.