InauguralSystems · InauguralPhysicist · Aug 2, 2026 · Aug 2, 2026
@@ -50,8 +50,11 @@ iteration, or a collector that quietly stops working).
   `make asan-http` (ext_http + model under ASan/UBSan; CI runs the suite that
   way, and the HTTP sections are probe-gated so they pull in automatically).
   Same for `ext_gfx.c` — in **no** default build; compile-check with
-  `make gfx`. All variants land on `src/eigenscript`, so never rebuild one
-  while a suite run against another is in flight.
+  `make gfx`. Variants coexist in per-variant `build/<variant>/` objdirs
+  (#740); `src/eigenscript` is a hard link to the last-built one, so a
+  rebuild no longer destroys another variant's binary — but running any
+  `make` variant target mid-suite still re-points the alias under the
+  suite (the #681 guard catches it).
 - **A per-request leak in `ext_http.c` will not be caught by any sanitizer
   gate.** LeakSanitizer runs atexit, and the server is torn down with `kill`
   against no SIGTERM handler, so LSan never runs in the server process —

@@ -4,6 +4,36 @@ All notable changes to EigenScript are documented here.
 
 ## [Unreleased]
 
+### Changed
+
+- **Build variants now coexist: per-variant objdirs with dependency
+  tracking (#740).** Every runtime variant (`build`/`full`/`http`/`zlib`/
+  `net`/`gfx`/`asan`/`asan-http`/`tsan`/`valgrind`/`poison`) was a single
+  whole-program `gcc` invocation writing `src/eigenscript`, so an ASan
+  build destroyed the release binary, any edit cost a full 22-TU rebuild,
+  and the project paid for the collision twice over — the test runner's
+  binary-fingerprint guard (#681) and a standing CLAUDE.md prohibition.
+  Each variant now compiles into its own `build/<variant>/` objdir with
+  `-MMD`/`-MP` header-dependency tracking and links
+  `build/<variant>/eigenscript`; `src/eigenscript` becomes a **hard
+  link** the phony target re-points, so every existing consumer of that
+  path works unchanged — hard rather than symbolic because the runtime
+  resolves the stdlib relative to `/proc/self/exe`, which dereferences a
+  symlink into `build/` and loses `lib/` (a symlink prototype failed 58
+  suite checks exactly there; the hard link keeps the exec'd path in
+  `src/`). Measured: switching `make asan` → `make` went from a ~90 s
+  full rebuild to a 0.2 s relink; touching `vm.c` recompiles one TU;
+  touching `vm.h` recompiles exactly its 12 includers. Target names,
+  flags per variant, and output messages are unchanged. `pgo`/`coverage`
+  (whole-program by nature) and `build.sh` now `rm -f` the alias first
+  so they write a fresh regular file rather than truncating the shared
+  inode under a variant. The fingerprint guard stays: re-pointing the
+  alias or relinking the same variant mid-suite is still a mid-run swap,
+  and the guard is what catches it (the guard's stat now uses `-L` so
+  its metadata half describes the same file its cksum half reads) — what
+  is retired is the cross-variant half of the prohibition, not the
+  guard.
+
 ### Fixed
 
 - **The `ValType` switches are now exhaustive too, closing out the #738

@@ -31,9 +31,9 @@ ours onto the new main. Rebasing our own costs nothing; asking them does. (Only
 ## Build & test
 
 ```
-make            # release build -> src/eigenscript (HTTP/MODEL/DB off)
+make            # release build -> build/release/, src/eigenscript hard-links to it (HTTP/MODEL/DB off)
 make test       # build + full suite (tests/run_all_tests.sh)
-make asan       # ASan+UBSan build (same binary path!) — extensions OFF
+make asan       # ASan+UBSan build — extensions OFF
 make asan-http  # ASan+UBSan *with* ext_http+model (CI gate; leaks still need RSS, see #731)
 make http       # http+model variant — run tests/test_http_server.sh
 make zlib       # DEFLATE codecs (inflate/deflate builtins) via system zlib (-lz)
@@ -69,7 +69,14 @@ bash tools/embed_stack_soak.sh  # embed REPL soak inside a 64 KiB stack rlimit (
   transition, and the JIT counters / OSR / inline-cache writes / trace-line are
   gated off under MT, name hashes precomputed at compile time. ThreadSanitizer
   here needs `setarch -R` to disable ASLR.)
-- `make asan` overwrites `src/eigenscript` — rebuild with `make` before timing.
+- Variants build into per-variant `build/<variant>/` objdirs (#740) and
+  coexist; `src/eigenscript` is a hard link to the last `make` target
+  (hard, not symbolic — `/proc/self/exe`-relative stdlib resolution must
+  keep seeing `src/`), so switching variants is an instant relink (`make`
+  after `make asan` costs ~0.2s, not a rebuild). Don't run any `make`
+  variant target while a suite is in flight — it re-points the alias
+  under the suite (the #681 fingerprint guard catches it at the next
+  section seam).
 - Benchmarks: `tests/bench_perf.eigs` (micro), `tests/bench_dmg_shape.eigs`
   (dispatch-table interpreter shape, the DMG/cpu_instrs stand-in),
   `tests/bench_idxset.eigs` (fn-local buffer/list write loop — one JIT thunk,

@@ -46,7 +46,95 @@ LSP_BINARY  := $(SRC_DIR)/eigenlsp
 DAP_SOURCES := $(SRC_DIR)/eigsdap.c $(SRC_DIR)/tape_read.c $(filter-out $(CLI_ONLY),$(SOURCES))
 DAP_BINARY  := $(SRC_DIR)/eigsdap
 
-.PHONY: all build full http net gfx zlib lib amalgamation tsan test install install-gfx clean coverage coverage-clean fuzz fuzz-run lsp dap jit-smoke embed-smoke asan valgrind pgo freestanding-check freestanding-libc-diff asan-http print-%
+.PHONY: all build full http net gfx zlib lib amalgamation tsan test install install-gfx clean coverage coverage-clean fuzz fuzz-run lsp dap jit-smoke embed-smoke asan valgrind pgo poison freestanding-check freestanding-libc-diff asan-http print-%
+
+# ---- Per-variant objdir engine (#740) -------------------------------------
+# The engine's rules are defined before `all`, so pin the default goal.
+.DEFAULT_GOAL := all
+# Every runtime variant compiles into its own build/<variant>/ objdir with
+# -MMD/-MP header-dependency tracking, links build/<variant>/eigenscript,
+# and the phony target re-points src/eigenscript at it (hard link — see
+# RELINK below for why not a symlink). So: variants COEXIST (make asan no
+# longer destroys the release binary — and switching back is an instant
+# relink, not a 22-TU rebuild), and rebuilds within a variant are
+# incremental. The alias keeps every existing consumer of src/eigenscript
+# working unchanged. The suite's fingerprint guard (#681) still applies:
+# re-pointing the alias or relinking the same variant mid-suite is caught
+# at the next section seam.
+VERDEF   := -DEIGENSCRIPT_VERSION='"$(VERSION)"'
+DEFS_OFF := -DEIGENSCRIPT_EXT_HTTP=0 -DEIGENSCRIPT_EXT_MODEL=0 -DEIGENSCRIPT_EXT_DB=0
+MODEL_SRC := $(SRC_DIR)/model_io.c $(SRC_DIR)/model_infer.c $(SRC_DIR)/model_train.c
+ASAN_FLAGS := -fsanitize=address,undefined -Werror=switch -g -O1
+
+SRC_V_release := $(SOURCES)
+FLAGS_release := $(CFLAGS) $(DEFS_OFF) $(VERDEF)
+LIBS_release  := $(LDFLAGS)
+
+SRC_V_full := $(FULL_SOURCES)
+FLAGS_full := $(CFLAGS) -I/usr/include/postgresql -DEIGENSCRIPT_EXT_NET=1 $(VERDEF)
+LIBS_full  := $(LDFLAGS) -lpq
+
+SRC_V_http := $(SOURCES) $(SRC_DIR)/ext_http.c $(MODEL_SRC)
+FLAGS_http := $(CFLAGS) -DEIGENSCRIPT_EXT_HTTP=1 -DEIGENSCRIPT_EXT_MODEL=1 -DEIGENSCRIPT_EXT_DB=0 $(VERDEF)
+LIBS_http  := $(LDFLAGS)
+
+SRC_V_zlib := $(SOURCES)
+FLAGS_zlib := $(CFLAGS) $(DEFS_OFF) -DEIGENSCRIPT_EXT_ZLIB=1 $(VERDEF)
+LIBS_zlib  := $(LDFLAGS) -lz
+
+SRC_V_net := $(SOURCES) $(SRC_DIR)/ext_net.c
+FLAGS_net := $(CFLAGS) $(DEFS_OFF) -DEIGENSCRIPT_EXT_NET=1 $(VERDEF)
+LIBS_net  := $(LDFLAGS)
+
+SRC_V_gfx := $(SOURCES) $(SRC_DIR)/ext_gfx.c
+FLAGS_gfx := $(CFLAGS) $(DEFS_OFF) -DEIGENSCRIPT_EXT_GFX=1 $(VERDEF)
+LIBS_gfx  := $(LDFLAGS) -ldl
+
+SRC_V_asan := $(SOURCES)
+FLAGS_asan := $(ASAN_FLAGS) $(DEFS_OFF) $(VERDEF)
+LIBS_asan  := -lm -lpthread
+
+SRC_V_asan-http := $(SOURCES) $(SRC_DIR)/ext_http.c $(SRC_DIR)/ext_net.c $(MODEL_SRC)
+FLAGS_asan-http := $(ASAN_FLAGS) -DEIGENSCRIPT_EXT_HTTP=1 -DEIGENSCRIPT_EXT_MODEL=1 -DEIGENSCRIPT_EXT_DB=0 -DEIGENSCRIPT_EXT_NET=1 $(VERDEF)
+LIBS_asan-http  := -lm -lpthread
+
+SRC_V_tsan := $(SOURCES)
+FLAGS_tsan := -fsanitize=thread -Werror=switch -g -O1 $(DEFS_OFF) $(VERDEF)
+LIBS_tsan  := -lm -lpthread
+
+SRC_V_valgrind := $(SOURCES)
+FLAGS_valgrind := -Werror=switch -g -O1 -DEIGS_VALGRIND $(DEFS_OFF) $(VERDEF)
+LIBS_valgrind  := -lm -lpthread
+
+SRC_V_poison := $(SOURCES)
+FLAGS_poison := -g -O1 -DEIGS_POISON $(DEFS_OFF) $(VERDEF)
+LIBS_poison  := -lm -lpthread
+
+VARIANTS := release full http zlib net gfx asan asan-http tsan valgrind poison
+
+# Objects depend on Makefile+VERSION so a flag or version-string change
+# rebuilds; header edits are covered by the generated .d files.
+define VARIANT_RULES
+OBJ_$(1) := $$(patsubst $(SRC_DIR)/%.c,build/$(1)/%.o,$$(SRC_V_$(1)))
+build/$(1)/%.o: $(SRC_DIR)/%.c Makefile VERSION | build/$(1)
+	$$(CC) $$(FLAGS_$(1)) -MMD -MP -c $$< -o $$@
+build/$(1)/eigenscript: $$(OBJ_$(1))
+	$$(CC) $$(FLAGS_$(1)) -o $$@ $$(OBJ_$(1)) $$(LIBS_$(1))
+build/$(1):
+	@mkdir -p $$@
+-include $$(OBJ_$(1):.o=.d)
+endef
+$(foreach V,$(VARIANTS),$(eval $(call VARIANT_RULES,$(V))))
+
+# The retarget lives in the phony targets below (not the link recipe) so
+# `make <variant>` always points src/eigenscript at that variant, even
+# when its binary was already up to date. HARD link, not symlink: the
+# runtime resolves lib/ relative to /proc/self/exe, which dereferences a
+# symlink to build/<variant>/ and would lose the executable-relative
+# stdlib; a hard link keeps the exec'd path at src/eigenscript.
+define RELINK
+@ln -f build/$(1)/eigenscript $(BINARY)
+endef
 
 # Introspection helper: `make print-SOURCES` echoes a variable's value.
 # tests/test_leak_guard.sh derives its ASan build source list from the
@@ -57,71 +145,37 @@ print-%:
 
 all: build
 
-build:
-	$(CC) $(CFLAGS) -o $(BINARY) $(SOURCES) \
-		-DEIGENSCRIPT_EXT_HTTP=0 \
-		-DEIGENSCRIPT_EXT_MODEL=0 \
-		-DEIGENSCRIPT_EXT_DB=0 \
-		-DEIGENSCRIPT_VERSION='"$(VERSION)"' \
-		$(LDFLAGS)
-	@echo "EigenScript $(VERSION) built. Binary: $$(du -sh $(BINARY) | cut -f1)"
+build: build/release/eigenscript
+	$(call RELINK,release)
+	@echo "EigenScript $(VERSION) built. Binary: $$(du -sh build/release/eigenscript | cut -f1)"
 
-full:
-	$(CC) $(CFLAGS) -o $(BINARY) $(FULL_SOURCES) \
-		-I/usr/include/postgresql \
-		-DEIGENSCRIPT_EXT_NET=1 \
-		-DEIGENSCRIPT_VERSION='"$(VERSION)"' \
-		$(LDFLAGS) -lpq
-	@echo "EigenScript $(VERSION) (full) built. Binary: $$(du -sh $(BINARY) | cut -f1)"
+full: build/full/eigenscript
+	$(call RELINK,full)
+	@echo "EigenScript $(VERSION) (full) built. Binary: $$(du -sh build/full/eigenscript | cut -f1)"
 
 # Build with HTTP + model extensions but without DB (no libpq-dev required).
 # Useful for running HTTP test suites on systems without PostgreSQL headers.
-http:
-	$(CC) $(CFLAGS) -o $(BINARY) $(SOURCES) \
-		$(SRC_DIR)/ext_http.c \
-		$(SRC_DIR)/model_io.c $(SRC_DIR)/model_infer.c $(SRC_DIR)/model_train.c \
-		-DEIGENSCRIPT_EXT_HTTP=1 \
-		-DEIGENSCRIPT_EXT_MODEL=1 \
-		-DEIGENSCRIPT_EXT_DB=0 \
-		-DEIGENSCRIPT_VERSION='"$(VERSION)"' \
-		$(LDFLAGS)
-	@echo "EigenScript $(VERSION) (http+model, no db) built. Binary: $$(du -sh $(BINARY) | cut -f1)"
+http: build/http/eigenscript
+	$(call RELINK,http)
+	@echo "EigenScript $(VERSION) (http+model, no db) built. Binary: $$(du -sh build/http/eigenscript | cut -f1)"
 
 # Build with the DEFLATE codecs (inflate/deflate builtins, #684) linked
 # against the system zlib. Same opt-in pattern as `make http`: the
 # default build stays zero-dependency and the four builtins raise
 # "compiled without zlib support" there.
-zlib:
-	$(CC) $(CFLAGS) -o $(BINARY) $(SOURCES) \
-		-DEIGENSCRIPT_EXT_HTTP=0 \
-		-DEIGENSCRIPT_EXT_MODEL=0 \
-		-DEIGENSCRIPT_EXT_DB=0 \
-		-DEIGENSCRIPT_EXT_ZLIB=1 \
-		-DEIGENSCRIPT_VERSION='"$(VERSION)"' \
-		$(LDFLAGS) -lz
-	@echo "EigenScript $(VERSION) (zlib) built. Binary: $$(du -sh $(BINARY) | cut -f1)"
+zlib: build/zlib/eigenscript
+	$(call RELINK,zlib)
+	@echo "EigenScript $(VERSION) (zlib) built. Binary: $$(du -sh build/zlib/eigenscript | cut -f1)"
 
 # Raw TCP sockets on the trace tape (#414). Same opt-in pattern as gfx:
 # in no default build, no extra library needed (plain POSIX sockets).
-net:
-	$(CC) $(CFLAGS) -o $(BINARY) $(SOURCES) $(SRC_DIR)/ext_net.c \
-		-DEIGENSCRIPT_EXT_HTTP=0 \
-		-DEIGENSCRIPT_EXT_MODEL=0 \
-		-DEIGENSCRIPT_EXT_DB=0 \
-		-DEIGENSCRIPT_EXT_NET=1 \
-		-DEIGENSCRIPT_VERSION='"$(VERSION)"' \
-		$(LDFLAGS)
-	@echo "EigenScript $(VERSION) (net) built. Binary: $$(du -sh $(BINARY) | cut -f1)"
+net: build/net/eigenscript
+	$(call RELINK,net)
+	@echo "EigenScript $(VERSION) (net) built. Binary: $$(du -sh build/net/eigenscript | cut -f1)"
 
-gfx:
-	$(CC) $(CFLAGS) -o $(BINARY) $(SOURCES) $(SRC_DIR)/ext_gfx.c \
-		-DEIGENSCRIPT_EXT_HTTP=0 \
-		-DEIGENSCRIPT_EXT_MODEL=0 \
-		-DEIGENSCRIPT_EXT_DB=0 \
-		-DEIGENSCRIPT_EXT_GFX=1 \
-		-DEIGENSCRIPT_VERSION='"$(VERSION)"' \
-		$(LDFLAGS) -ldl
-	@echo "EigenScript $(VERSION) (gfx) built. Binary: $$(du -sh $(BINARY) | cut -f1)"
+gfx: build/gfx/eigenscript
+	$(call RELINK,gfx)
+	@echo "EigenScript $(VERSION) (gfx) built. Binary: $$(du -sh build/gfx/eigenscript | cut -f1)"
 
 test: build
 	cd tests && bash run_all_tests.sh
@@ -212,13 +266,8 @@ embed-smoke: amalgamation
 # the normal -O2 build silently tolerates. ~2x slower; for testing only.
 # The full suite runs leak-clean, so leave leak detection on:
 #   make asan && cd tests && ASAN_OPTIONS=detect_leaks=1 bash run_all_tests.sh
-asan:
-	$(CC) -fsanitize=address,undefined -Werror=switch -g -O1 -o $(BINARY) $(SOURCES) \
-		-DEIGENSCRIPT_EXT_HTTP=0 \
-		-DEIGENSCRIPT_EXT_MODEL=0 \
-		-DEIGENSCRIPT_EXT_DB=0 \
-		-DEIGENSCRIPT_VERSION='"$(VERSION)"' \
-		-lm -lpthread
+asan: build/asan/eigenscript
+	$(call RELINK,asan)
 	@echo "EigenScript $(VERSION) (asan+ubsan) built. Binary: $(BINARY)"
 
 # ASan+UBSan over the EXTENSION surface — same variant as `make http`
@@ -233,43 +282,24 @@ asan:
 # would make this unbuildable on a machine without postgres. ext_db.c
 # therefore remains unsanitized — a separate, smaller gap.
 #   make asan-http && cd tests && ASAN_OPTIONS=detect_leaks=1 bash run_all_tests.sh
-asan-http:
-	$(CC) -fsanitize=address,undefined -Werror=switch -g -O1 -o $(BINARY) $(SOURCES) \
-		$(SRC_DIR)/ext_http.c $(SRC_DIR)/ext_net.c \
-		$(SRC_DIR)/model_io.c $(SRC_DIR)/model_infer.c $(SRC_DIR)/model_train.c \
-		-DEIGENSCRIPT_EXT_HTTP=1 \
-		-DEIGENSCRIPT_EXT_MODEL=1 \
-		-DEIGENSCRIPT_EXT_DB=0 \
-		-DEIGENSCRIPT_EXT_NET=1 \
-		-DEIGENSCRIPT_VERSION='"$(VERSION)"' \
-		-lm -lpthread
+asan-http: build/asan-http/eigenscript
+	$(call RELINK,asan-http)
 	@echo "EigenScript $(VERSION) (asan+ubsan, http+model+net) built. Binary: $(BINARY)"
 
 # ThreadSanitizer build for the concurrency race gate (tests/test_tsan.sh).
 # Complements ASan (which is not run with the thread checker). Run the tests
 # under `setarch -R` — ThreadSanitizer needs ASLR disabled here (CLAUDE.md).
-tsan:
-	$(CC) -fsanitize=thread -Werror=switch -g -O1 -o $(BINARY) $(SOURCES) \
-		-DEIGENSCRIPT_EXT_HTTP=0 \
-		-DEIGENSCRIPT_EXT_MODEL=0 \
-		-DEIGENSCRIPT_EXT_DB=0 \
-		-DEIGENSCRIPT_VERSION='"$(VERSION)"' \
-		-lm -lpthread
+tsan: build/tsan/eigenscript
+	$(call RELINK,tsan)
 	@echo "EigenScript $(VERSION) (tsan) built. Binary: $(BINARY)"
 
 # Plain -O1 -g minimal build for Valgrind/Memcheck (tests/valgrind_smoke.sh).
 # No sanitizers — Valgrind shadows the uninstrumented binary at runtime, so it
 # complements ASan/UBSan/TSan (uninit reads, UAF, definite/indirect leaks) on a
 # system without instrumented libs. -O1 keeps optimizer-induced false positives
 # down while giving usable stacks. Same minimal extension surface as asan.
-valgrind:
-	$(CC) -Werror=switch -g -O1 -o $(BINARY) $(SOURCES) \
-		-DEIGS_VALGRIND \
-		-DEIGENSCRIPT_EXT_HTTP=0 \
-		-DEIGENSCRIPT_EXT_MODEL=0 \
-		-DEIGENSCRIPT_EXT_DB=0 \
-		-DEIGENSCRIPT_VERSION='"$(VERSION)"' \
-		-lm -lpthread
+valgrind: build/valgrind/eigenscript
+	$(call RELINK,valgrind)
 	@echo "EigenScript $(VERSION) (valgrind -O1 -g) built. Binary: $(BINARY)"
 
 # Uninitialized-read hunter (the EigenOS #UD heisenbug class, see
@@ -279,14 +309,8 @@ valgrind:
 # benign zero pages. Run the suite against it, with the raw-malloc boundary
 # poisoned too:
 #   make poison && cd tests && MALLOC_PERTURB_=170 bash run_all_tests.sh
-poison:
-	$(CC) -g -O1 -o $(BINARY) $(SOURCES) \
-		-DEIGS_POISON \
-		-DEIGENSCRIPT_EXT_HTTP=0 \
-		-DEIGENSCRIPT_EXT_MODEL=0 \
-		-DEIGENSCRIPT_EXT_DB=0 \
-		-DEIGENSCRIPT_VERSION='"$(VERSION)"' \
-		-lm -lpthread
+poison: build/poison/eigenscript
+	$(call RELINK,poison)
 	@echo "EigenScript $(VERSION) (poison 0xAA -O1 -g) built. Binary: $(BINARY)"
 
 # Profile-guided optimization. Builds an instrumented binary, runs the
@@ -298,6 +322,7 @@ PGO_DIR ?= /tmp/eigs-pgo
 PGO_RUN ?= cd $(HOME)/DMG && $(CURDIR)/$(BINARY) dmg.eigs roms/cpu_instrs.gb --cycles 200000 >/dev/null
 pgo:
 	@rm -rf $(PGO_DIR) && mkdir -p $(PGO_DIR)
+	@rm -f $(BINARY)   # may be a variant symlink — never write through it
 	$(CC) $(CFLAGS) -fprofile-generate=$(PGO_DIR) -o $(BINARY) $(SOURCES) \
 		-DEIGENSCRIPT_EXT_HTTP=0 \
 		-DEIGENSCRIPT_EXT_MODEL=0 \
@@ -330,6 +355,7 @@ coverage: coverage-clean
 			-DEIGENSCRIPT_EXT_DB=0 \
 			-DEIGENSCRIPT_VERSION='"$(VERSION)"' || exit 1; \
 	done
+	@rm -f $(BINARY)   # may be a variant symlink — never write through it
 	$(CC) --coverage -o $(BINARY) $(SOURCES:.c=.o) $(LDFLAGS)
 	-cd tests && bash run_all_tests.sh > /dev/null 2>&1 || true
 	@cd $(SRC_DIR) && gcov -n -b $(notdir $(SOURCES)) > ../coverage.txt 2>&1 || true

@@ -31,6 +31,12 @@ fi
 # JIT_FLAGS=-DEIGENSCRIPT_JIT_FORCE_OFF=1`).
 JIT_FLAGS=""
 
+# The Makefile's objdir engine (#740) leaves src/eigenscript as a hard link
+# to build/<variant>/eigenscript. Remove the name before compiling so
+# build.sh writes a fresh file instead of truncating the shared inode under
+# that variant's binary.
+rm -f eigenscript
+
 if [ "$1" = "lsp" ]; then
     # Language server (src/eigenlsp) — the editor-intelligence half of the
     # toolchain. Links eigenlsp.c against the runtime (SOURCES minus the