v1.10.0 — heal→MAPE-K, JIT IR opt, web-native addressed LM + 45% lossless web compression

Core (verified — 38/38 core tests green)

• Autograd tape numerical-stability sweep. ε-clamp tape_log, overflow-clamp tape_exp, and fix the tape_pow_int backward 0·∞→NaN poison. Full sweep: div/softmax/mean/sigmoid verified safe.
• Self-healer → MAPE-K. Persistent cross-pass Knowledge (HealClassCounts.add / last_fixpoint_knowledge) + set-signature convergence (vs bare count-equality), surfaced as heal-summary in --check and OMC_HEAL. HEAL validated as a generator-repair operator.
• Safe self-modification. fn_swap_verified gains a determinism gate + an invariant SET (array 3rd arg — accept only if every invariant holds).
• JIT LLVM-IR optimization. omnimcode-codegen optimize() runs a curated scalar pass pipeline (mem2reg / instcombine / reassociate / gvn / simplifycfg), LSR deliberately excluded (known segfault on dual-band LCSSA loops), env-gated in the dispatch builder (OMC_HBIT_JIT_OPT).
• Addressing. fns_on_subface (3-level HAddr descent); memo cache made FIFO-bounded.

Research — web-native addressed LM (experiments/transformerless_lm/)

A transformerless LM built entirely on the addressed knowledge web — no token-prediction model anywhere.

• Two empirical oracles (held-out): execution/resolve (a sentence executes by traversing its concept-addresses over hub-damped PMI edges; AUC 0.91–0.98 real-vs-salad) and fluency (learned from the web's own transitions; AUC 0.86).
• heal-to-resolution, a concepts→sentence realizer, a recall|relate|decline tool-agent, create-by-recombination across distant sources (3-gated), and a write-don't-train self-improvement loop.
• Lossless 45% knowledge-web compression — the LM was asked how to compress itself; node-string interning + zlib delivered 9.78 GB → 5.33 GB, verified reversible, presented to all readers via SQLite views (zero reader changes).

See experiments/transformerless_lm/README.md. Data (knowledge.db etc.) is regenerable and not shipped in-repo.