build: one host builds every platform; PLATFORM selects, dist ships all - #654

Four follow-ups from review, all in one commit.

1. PLATFORM selects; the defaults differ by target

make build                                   # this machine
make build PLATFORM=linux/amd64
make build PLATFORM=all                      # every supported platform
make dist                                    # every supported platform
make dist  PLATFORM="linux/amd64 darwin/arm64"

ALL_PLATFORMS names what the project supports. PLATFORM is the caller's selection and stays empty
unless named, because you build what you are about to run, and you ship everything — opposite
defaults, so the selection resolves per target through $(call select,<default>) rather than through a
variable with one default.

Measured, warm: 16.5s for the host against 47.5s for the full matrix. Full-matrix-by-default was
a ~3x tax on every local build, and the relief valve only ever helped developers who remembered it.
Cross-compilation is now one word away instead.

override turned out to be required, not stylistic: PLATFORM=all arrives on the command line, and a
command-line variable wins over any plain makefile assignment. The first attempt reassigned it inside an
ifeq, which make silently ignored — producing build/all-all/ and a failed build. Caught by running
it; recorded in the comment so the next person does not rediscover it.

2. devlore-test is not a shipping product

Moved from PRODUCTS to TOOLS. It is the graph test harness and it does not ship today, so
cross-compiling it produced binaries nobody can run and nothing distributes. build/<platform>/ now
holds exactly lore, star, writ.

Checked before moving it: the self-install scenario resolves lore, star, writ — not
devlore-test.

3. Flat host copies for lore and writ

New HOST_COPIES. Products that also get run from this checkout keep a flat build/<name> beside their
per-platform copy, so "the binary that runs on this machine" has one path that does not shift with the
host triple — the guarantee $(STAR) has always relied on for build/star.

star is deliberately absent: build/star comes from the star: target via the build/star: FORCE
bootstrap, which must run before codegen and therefore before this list is reached. Adding it would
build identical bytes twice and put two recipes in charge of one path.

Verified byte-identical to their host-platform products.

4. One make run, one stamp; host tools are static

• CGO_ENABLED=0 folded into HOST_GO. The matrix built products with it while the star: target
  inherited the ambient default — cgo on. build/star and build/<host>/star were genuinely
  different binaries, one linking the cgo implementations of net and os/user and one the pure-Go
  ones. That was introduced by this PR's first commit; folding the flag into HOST_GO fixes it for
  every host tool at once.

• export VERSION COMMIT BUILD_DATE. build/star: FORCE shells out to $(MAKE) star, and a
  sub-make re-parses this file — a second date, a second git describe. The two binaries agreed
  locally only because a warm Go cache kept both inside one tick of date; a cold cache or a loaded
  runner separates them. Exporting puts the frozen values in the sub-make's environment, where the ?=
  guards find them already defined.

  This extends the simply-expanded fix at Makefile:44-51 rather than reversing it. That fix stopped
  re-evaluation within a run — the Windows runner comparing 8a62b17-dirty against a binary reporting
  8a62b17. The sub-make is the one boundary it did not reach, and the freeze comment already describes
  the mechanism used here: "?= above still honors an override from the environment."

  Verified by mechanism, not coincidence: with BUILD_DATE=2000-01-01T00:00:00Z in the environment, the
  sub-make's build/star and the cross-compiled products all report that value, read out of the linker
  flags with go version -m.

5. Two recursive-make hacks removed

test-scenario and docs had been narrowed with $(MAKE) build PLATFORM=<host>. Host-only is now the
default, so both are plain : build prerequisites again.

Verification

make build default / PLATFORM=all / PLATFORM=linux/amd64 / PLATFORM="linux/amd64 darwin/arm64" —
all 0, each producing exactly the selected directories, with the stamp-check skip announced when the host
is excluded. make dist 0 (six archives), make dist PLATFORM=linux/amd64 0 (one). make test 103 ok ·
make vet darwin/linux/windows · test-scenario · docs · complexity · shell-lint ·
verify-ldflags · gofmt -l — all clean.

Correction to the follow-up comment above: it says "override turned out to be required, not stylistic." There is no override in the merged design, and none is needed.

That sentence describes a discarded intermediate. The first attempt made PLATFORM a variable with a default and tried to expand all by reassigning it inside an ifeq — which make silently ignores for a command-line variable, producing build/all-all/ and a failed build. override would have fixed that design.

The design that landed does not reassign PLATFORM at all. select reads it at expansion time and filters all there:

PLATFORM ?=
select = $(if $(PLATFORM),$(if $(filter all,$(PLATFORM)),$(ALL_PLATFORMS),$(PLATFORM)),$(1))

A function was needed anyway, because build and dist want opposite defaults and one variable cannot hold two. Sidestepping the command-line-precedence problem is a consequence of that, not a separate fix. The Makefile comment at the select definition states this correctly; only the PR comment was wrong.