feat: context sliding window (infinite generation beyond max_seq_len)

[cstroie]

Summary

• Implements KV cache sliding window so generation continues past max_seq_len without truncation
• When pos - kv_shift >= max_seq_len, kvcache_slide() evicts half the non-prefix entries via memmove across all layers, then increments kv_shift; the physical slot is always logical_pos - kv_shift
• Adds -w <int> flag (keep_prefix) to pin BOS / system-prompt tokens at the front of the cache so they are never evicted

Compliance with CONTRIBUTING.md

Rule	Status
Zero dependencies (libc/libm/libpthread only)	✅ kvcache_slide uses only memmove
No malloc during inference	✅ operates entirely on pre-allocated KV cache buffers
Works on 256 MB RAM	✅ memory footprint unchanged — same fixed KV cache
Plain C11, no C++	✅
snake_case / 4-space indent / type_t suffix	✅

Verification output (x86-64, AVX2, TinyLlama 1.1B Q4_K_M)

# make clean && make native
gcc -O3 -std=c11 -D_GNU_SOURCE -Wall -Wextra -Wpedantic -march=native -o picolm ...

# greedy test
$ ./picolm model.gguf -p "The capital of France is" -n 20 -t 0
 Paris.
2. B.C. The capital of ancient Rome was Rome.
3...

# JSON mode
$ ./picolm model.gguf --json -p "Return JSON with a name" -n 50 -t 0.3
[]

# memory check
$ ./picolm model.gguf -p "Hello" -n 10 2>&1 | grep Memory
Memory: 45.17 MB runtime state (FP16 KV cache)

Hardware tested

• x86-64 Linux, AVX2 (Intel/AMD Haswell+)

🤖 Generated with Claude Code