cleanup: retire the legacy config surface — env seeding, turboquant aliases, dead flags

[kekzl]

What

Simplification pass over the config/CLI layer, follow-up to #878's consistency audit. −305/+70 LOC. Every removal was verified to have zero producers/consumers in the repo (grep across src/, tools/, tests/, scripts/, .github/, Dockerfile, docker-entrypoint); all removed knobs stay reachable via imp.conf / --set.

Changes

Legacy env seeding trimmed 29 → 2 (seed_from_env, src/runtime/config.cpp): only IMP_DETERMINISTIC (set by test_determinism_e2e/test_lora to inject determinism through the public API) and IMP_FMHA_FA2 (set by the roofline A/B harness per subprocess) have real producers — the other 27 vars had none anywhere. All stale Legacy env: IMP_* comments and log strings swept repo-wide (config.h, kernels, resolver logs); CLAUDE.md and imp.conf.example updated to match.

--kv-fp16 was a silent no-op — now fixed: it setenv'd IMP_KV_FP16, which no code has ever read (the comment claiming "the engine reads it" was wrong). It now pushes kv_cache.dtype=fp16, actually forcing FP16 KV and opting out of the auto FP8-hint upgrade — matching its documented semantics. --no-fp8-prefill similarly drops its env round-trip for a direct attention.fp8_prefill=never override.

TurboQuant surface fully retired (deprecated since Phase 5, 2026-05-17; zero callers): --kv-turboquant{,-lite} + --tq-sketch-mult flags in both binaries, the per-request JSON overrides, the one-shot WARN blocks duplicated in imp-cli/main.cpp and imp-server/handlers.cpp, the IMP_DTYPE_TURBOQUANT(_LITE) C-API enum aliases, and the ignored turboquant_sketch_multiplier ImpConfig field. (The live quant/turboquant_fp4.cuh FP4 helpers used by MXFP4-KV are untouched.)

Dead code deleted: --cuda-graphs no-op flag; ImpConfig.gpu_memory_pool_size and ImpConfig.num_cpu_threads (default-initialized, never read anywhere).

Deliberately NOT touched (audited, kept)

• Default-off research scaffolds from closed programs (mxfp4 attention family, BitDecoding, FFN sparsity, q4k_hmma, spec_capture_probe) — kept per the explicit idea(attention): NVFP4 attention compute on sm_120 — SageAttention3 recipe (K-smoothing + per-16-block scales + two-level P) #846/release: v0.16.2 — FP4-attention research batch (#846 program closed) #873 decision to ship them as quality-validated scaffolds.
• Dev-only getenv(IMP_SPEC_TRACE/IMP_JUMP_TRACE/IMP_PPL_DUMP) trace toggles.
• The two near-duplicate args parsers — 27 byte-identical branches, but unifying them is speculative abstraction without a concrete bug; the ugliest duplication (turboquant WARN blocks) dies in this PR anyway.

Verification

• make build + make test-unit (37/37) + make verify-fast (fast gtest filter PASS).
• GPU smoke (Qwen3-8B-Q8_0, both arms): --kv-fp16 --no-fp8-prefill resolves end-to-end (log: FP8 prefill: disabled (attention.fp8_prefill=never)), generation coherent and identical to the default arm.
• Removed flags fail fast: imp-cli --kv-turboquant / imp-server --tq-sketch-mult → Unknown argument.
• No hot-path change (kernel files only had comment/log-string edits); baseline untouched.

🤖 Generated with Claude Code