fix(lookup): free charset_len_buffer (GPU buffer leak aborting multi-config-group lookups)

[bandrel]

Symptom

Multi-config-group lookups abort partway through with Failed to create context: -6 (CL_OUT_OF_HOST_MEMORY) — observed at config group ~7 of a full-size table campaign.

Root cause

charset_len_buffer is created in both host_thread_false_alarm() and host_thread_precompute() but freed in neither — only the batch-precompute path (batch_precompute_all_hashes) released it. Every GPU false-alarm / per-hash-precompute invocation leaked one GPU memory buffer. Across config groups these accumulate on the device until clCreateContext() can no longer allocate.

Evidence

Wrapped the OpenCL create/release entry points with live counters and dumped them per config group:

group-1-start  ctx=0 buf=0  ...
group-2-start  ctx=0 buf=6      <- +6 buffers, contexts/queues/programs/kernels all balanced
group-3-start  ctx=0 buf=12
...
group-10-start ctx=0 buf=54

Only buf climbed (monotonic, ~one per GPU false-alarm invocation). After adding the frees, buf stays flat at 0 across all 10 groups and the full run completes (exit=0, 0 context errors, potential-match counts unchanged).

Fix

Add the missing CLFREEBUFFER(charset_len_buffer) to both cleanup paths. Shared host code, so this affected all backends (OpenCL/CUDA/Metal).

Verification

• make linux, make cuda, make macos all build.
• dell3 OpenCL run: buffer counter flat across 10 config groups (was +N/group); full lookup completes cleanly.

🤖 Generated with Claude Code