feat(engine): replace the ablation scoring mode with a checked, digest-bound one

.github/workflows/release.yml

@@ -136,6 +136,11 @@ jobs:
           # run it -- the second half of #720.
           mkdir -p dist/tools
           cp c/tools/k3_tokenizer.py dist/tools/
+          # check_ablate_evidence.py verifies the ABLATE evidence that eval_glm.py
+          # (packaged above) produces. Nothing in the code names it -- its caller is a
+          # person with argparse arguments -- so pack_python's import walk cannot reach
+          # it, exactly like k3_tokenizer.py. Copied explicitly for the same reason.
+          cp c/tools/check_ablate_evidence.py dist/tools/
           cp c/coli dist/
           # Windows has nothing to click otherwise: `coli` is an extensionless
           # Python script, and the .exe files are engines that exit at once
@@ -190,6 +195,9 @@ jobs:
           # Kimi K3 needs tokenizer.json synthesized from tiktoken.model; without this
           # script in the archive the engine ships but cannot be driven by coli chat.
           test -f tools/k3_tokenizer.py || { echo "FAIL: k3_tokenizer.py missing from archive"; exit 1; }
+          # Same position as k3_tokenizer.py: reachable only by a person, so only an
+          # explicit gate can prove it shipped.
+          test -f tools/check_ablate_evidence.py || { echo "FAIL: check_ablate_evidence.py missing from archive"; exit 1; }
           # E tutto cio' che coli raggiunge, import e sottoprocessi. Il controllo
           # qui sopra guarda l'unico file che il passo di copia aveva appena
           # scritto: conferma se stesso, ed e' il motivo per cui la v1.10.0 e'
@@ -202,6 +210,18 @@ jobs:
           # fine. The check was wrong, not the artifact.
           python3 -c "import ast; ast.parse(open('tools/k3_tokenizer.py', encoding='utf-8').read())" \
             || { echo "FAIL: packaged k3_tokenizer.py does not parse"; exit 1; }
+          python3 -c "import ast; ast.parse(open('tools/check_ablate_evidence.py', encoding='utf-8').read())" \
+            || { echo "FAIL: packaged check_ablate_evidence.py does not parse"; exit 1; }
+          # test -f and ast.parse both only look at this one file, and only at its
+          # syntax -- neither one executes it, so neither catches a real import-time
+          # failure: an undefined name at module scope, a decorator that raises, or
+          # (were a local import ever reintroduced here) a dependency pack_python's
+          # walk did not carry into the archive. Import it for real, from the archive
+          # root, the same way a person running the shipped `coli` would reach it --
+          # never execute it, an import is enough to prove every name it needs at
+          # module load resolves.
+          python3 -c "import tools.check_ablate_evidence" \
+            || { echo "FAIL: packaged check_ablate_evidence.py does not import"; exit 1; }
           COLI_EXPECT="$SIBLINGS" python3 - <<'PYCHK'
           import importlib.machinery, importlib.util, json, os, sys, tempfile
           from family_registry import all_families

c/Makefile

@@ -733,7 +733,7 @@ $(file >.build-config,$(BUILD_CONFIG))
 endif
 .build-config: ;
 
-colibri$(EXE): colibri.c cli_args.h st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h quant.h sample.h kv_persist.h telemetry.h route_trace.h omp_tune.h kv_fp8.h kv_tq.h abl.h backend_cuda.h backend_metal.h backend_vulkan.h decode_batch.h edge_adapters.h edge_runtime.h edge_tok_internal.h schema_gbnf.h segment_adapter_internal.h segment_adapters.h segment_runtime.h tier.h $(CUDA_OBJ) $(METAL_OBJ) $(VK_OBJ) $(VK_SPV) .build-config
+colibri$(EXE): colibri.c cli_args.h st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h quant.h sample.h kv_persist.h telemetry.h route_trace.h omp_tune.h kv_fp8.h kv_tq.h abl.h backend_cuda.h backend_metal.h backend_vulkan.h decode_batch.h edge_adapters.h edge_runtime.h edge_tok_internal.h evidence_digest.h schema_gbnf.h segment_adapter_internal.h segment_adapters.h segment_runtime.h tier.h $(CUDA_OBJ) $(METAL_OBJ) $(VK_OBJ) $(VK_SPV) .build-config
 	$(CC) $(CFLAGS) colibri.c $(CUDA_OBJ) $(METAL_OBJ) $(VK_OBJ) -o colibri$(EXE) $(LDFLAGS)
 
 # Vulkan backend object (plain C + vulkan headers) and its SPIR-V shaders.
@@ -1321,6 +1321,20 @@ tests/test_rope_invfreq$(EXE): tests/test_rope_invfreq.c colibri.c st.h uring.h
 tests/test_grammar_cache$(EXE): tests/test_grammar_cache.c colibri.c st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h quant.h sample.h kv_persist.h telemetry.h route_trace.h
 	$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)
 
+# The classified logit-row reduction in sample.h and the digest in
+# evidence_digest.h: no model/weights, published SHA-256 vectors, an
+# agreement check against the plain logprob_target the sampling path uses
+# on rows a float subtraction cannot mis-round, and a double-precision
+# check against a second double computation on rows that do.
+tests/test_logprob_status$(EXE): tests/test_logprob_status.c colibri.c sample.h evidence_digest.h st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h quant.h kv_persist.h telemetry.h route_trace.h
+	$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)
+
+# The ablation mode's manifest loader, evidence writer and dispatch contract.
+# The adapter build bypasses only the model computation, so the parser/writer
+# path is exercised for real with no model and no weights.
+tests/test_ablate_mode$(EXE): tests/test_ablate_mode.c colibri.c sample.h evidence_digest.h abl.h st.h uring.h json.h tok.h tok_unicode.h compat.h grammar.h tier.h quant.h kv_persist.h telemetry.h route_trace.h
+	$(CC) $(CFLAGS) $< -o $@ $(LDFLAGS)
+
 # Standalone: drives a faithful miniature of moe()'s routing+accumulate and links
 # the SAME abl.h the engine links -- no model/weights needed (the ablation-logic gate).
 tests/test_ablate$(EXE): tests/test_ablate.c abl.h