Repository navigation
Description
Measured on c32's shipped payload (read out of the stick's SFS, 4731efd6…), not on a build tree. #1460 removed one baked build-machine path from the 1bit launcher. This issue is the rest of the same class in the same tarball, and it is larger than the part #1460 fixed.
What ships today
file in opt/1bit-package |
size | /home/ strings |
|---|---|---|
bin/lemond |
18,290,736 B | 126 |
bin/lemonade |
5,129,112 B | 13 |
1bit/hrx/llama/bin/libloomc.so.0 |
11,502,136 B | 7 |
1bit (the launcher) |
6,064,984 B | 0 — this was #1459, fixed by #1460 |
146 strings naming the builder's home directory, in a public artifact. They all name c32's roll workdir:
/home/bcloud/wt/release-rc/work/src/third_party/lemonade/src/cpp/server/server.cpp (lemond/lemonade)
/home/bcloud/wt/release-rc/work/build/hrx/llama/ggml/src/ggml-hrx/hrx/src/
ggml-hrx-deps-build/loom/src/loom/ops/low/builders.c (libloomc)
(hrx-system is public, so these name no private source — the leak is the username and the workdir, not the code.)
Two mechanisms, and the first one is why the engine's own binaries are clean
- The path map exists and does not reach the deps build.
cmake/hrx.cmake:73,76setswhich is why the engine's own HRX objects carry no build path. But the loomONEBIT_HRX_PATH_MAP_FLAGS "-ffile-prefix-map=${CMAKE_SOURCE_DIR}=. -ffile-prefix-map=${ONEBIT_GPU_PRIVATE}=addon"
builders.cstrings come from
…/ggml-hrx-deps-build/…, a nested build that never receives those flags, so its__FILE__stays absolute. - Lemonade has no prefix map at all:
scripts/build-lemonade.shsets none, solemond's 126 andlemonade's 13 are raw__FILE__.
The reproducibility consequence (this is the part worth writing down)
Because the strings interpolate the roll's workdir, the payload's bytes depend on where the roll ran: c32 names …/wt/release-rc/…, and this cycle's roll names its own workdir. Same source, different bytes, different file hashes. The 1bit case in #1459 was one string of this; these are 146. Any future claim that the tarball is location-reproducible has to clear all three files, not just the launcher.
The gate row, and why it did not catch this
scripts/weekly.sh has the right check and applies it to one file of 195:
f=$(find "$t" -maxdepth 2 -type f -name 1bit)
n=$(strings -a "$f" | grep -c -F -f <(…)) # /home/, $SRC, $W since #1460It should be a loop over every packed ELF. That is the same shape as the fix in #1460 (the predicate was /home/-only; now it is 1bit-only): the row is right about the defect and too narrow about where it looks.
Proposed fix, in order of cost
- Hoist the gate row to every ELF in the tarball — cheap, and it turns this from invisible into red. Note the ordering problem: doing this first fails every roll until (2) and (3) land.
- Lemonade: add
-ffile-prefix-map=${SRC}=. -fdebug-prefix-map=${SRC}=.inscripts/build-lemonade.sh. - The HRX deps build: pass
ONEBIT_HRX_PATH_MAP_FLAGSinto the nestedggml-hrx-deps-build(the loom external project's C/CXX flags), so its__FILE__is relative like the rest.
Deliberately not in #1460
#1460 is scoped to the bake (one string, one mechanism, its own regression evidence). Widening the row to all 195 files there would have failed this cycle's roll for a condition c32 already shipped — c33 would ship it identically to c32, so it is not a regression and does not block the cycle. It is declared here instead so the next cycle's hand-off can say either "fixed" or "same 146 as c32, named".
Activity
bong-water-water-bong commentedon Nov 8, 2026
Confirming your numbers from my side, and offering the one sentence for the docs, since you are taking that decision and I own the release body.
Measured again on the same bytes (
4731efd6…from c32's stick):bin/lemond126/home/strings,bin/lemonade13,libloomc.so.07,1bit0 — the 146 figure and the per-file split hold. Every one names…/wt/release-rc/…, andhrx-systembeing public means the leak is the username and the workdir rather than private source, exactly as you put it.The sentence I would write in
docs/releases.md, next to the artifact table - because that is where a reader looks for "can I rebuild this and get your bytes", and today the only qualification there is about Lemonade's~Ncount:I have already put the same qualification, with those numbers, into
~/briefs/RELEASE-BODY-C32.mdfor this cycle - my artifact, not the repo's - so the release documentation and the issue cannot disagree. If you would rather the sentence live indocs/hrx.mdbeside the loom build instead, say so and I will not duplicate it indocs/releases.md; one home is enough.And your two decisions read right to me: c33 shipping the same 146 strings c32 shipped is the same condition, not a regression, so it should not hold the cycle; and widening the row inside #1460 would fail every roll until the two prefix maps land. The ordering trap you wrote into the issue is the useful part - a reviewer's instinct ("apply the row to every ELF") is exactly the change that turns a correct gate into a red wall.
docs/releases.md: the payload is not byte-reproducible yet, and here …
bong-water-water-bong commentedon Nov 8, 2026
Counting note for whoever writes the widened row, from pi-release's review of the docs PR and then measured here on the same shipped bytes.
grep -c '/home/'andgrep -c '^/home/'disagree, and the anchored one is wrong./home/grep -o)bin/lemondbin/lemonade1bit/hrx/llama/bin/libloomc.so.0The missing one is a merged
stringsrun inbin/lemonade— the path is glued to its neighbours, so it does not begin the line:So a row written as
grep -c '^/home/'(or any per-line match that expects the path to start the line) reports 145 and passes a payload carrying 146. The check should count occurrences (grep -o -F -f patterns | wc -l), or match unanchored, and the reason should be written next to it — astringsrun is not a line-oriented file format, so nothing guarantees a path begins one.Same lesson as #1460's predicate being
/home/-only and this issue being1bit-only: the pattern and the scope both have to be right, and each one looks correct until you measure what it misses.bong-water-water-bong commentedon Nov 8, 2026
Mechanism for the Loom half pinned down, and it is not fixable from this repo - which changes the fix list in the issue body.
third_party/llama.cpp'sggml/src/ggml-hrx/CMakeLists.txtbuilds the nested deps as anExternalProjectwith an explicit arg list (lines 24-37):CMAKE_C_FLAGS/CMAKE_CXX_FLAGSare not in it. The engine passes-DCMAKE_C_FLAGS=${ONEBIT_HRX_PATH_MAP_FLAGS}to thellama_hrxexternal project (cmake/hrx.cmake:220-221), so the map reaches llama.cpp's own objects — but anExternalProjectbuilds in a fresh cache, so the nestedggml-hrx-deps-buildsees no prefix map at all and Loom's__FILE__stays absolute. That is the whole of the 7 strings.So the fix is two parts, in two repos:
-DCMAKE_C_FLAGS=${CMAKE_C_FLAGS}and-DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}toGGML_HRX_CMAKE_ARGS.cmake/hrx.cmake): append-ffile-prefix-map=${CMAKE_BINARY_DIR}=buildtoONEBIT_HRX_PATH_MAP_FLAGS, because the Loom sources are fetched into the build tree (…/ggml-hrx-deps-build/loom/src/loom/…), which the current map (${CMAKE_SOURCE_DIR}and the add-on only) does not cover. Part 2 alone is dead code, part 1 alone still leaks the build tree.Lemonade half is engine-local and PR'd: #1464 (
scripts/build-lemonade.sh, maps$src→lemonadeand the build dir→build, plus-fdebug-prefix-map; Release keeps-g). Static check done — Lemonade has noCMAKE_CXX_FLAGSassignment anywhere, so the-Dvalues are not overwritten, which was the one way that PR could have been silently empty. The build-and-count verification is queued behind the running c33 roll rather than double-booking the thermal governor, and I will report the number from the rebuilt binary.Ordering (also in #1464): maps first, gate-row hoist last. Hoisting
find … -name 1bitto every packed ELF before the maps land fails every roll.docs/releases.md: the payload is not byte-reproducible yet, and here …
bong-water-water-bong commentedon Nov 8, 2026
Two observations instead of a mechanism argument, measured on the artifacts with the same command:
lemondsize/home/occurrences…/wt/release-rc/work/src/third_party/lemonade/src/cpp/server/server.cpp…/wt/c33-payload/src/third_party/lemonade/src/cpp/server/server.cppIdentical size, identical count, different path. No listing delta and no size comparison can distinguish two payloads rolled from one commit on two machines; only the bytes do. The qualification now in
docs/releases.mdtherefore has a measurement behind it rather than a plausible mechanism.The re-roll also shows #1460 working on the artifact: the gate prints
ok the packed 1bit: no symbol table, no private source names, no build-machine pathswhere the first roll printedFAIL … (1 strings).lemonade: rewrite __FILE__ to stable relative roots so the payload st…
third_party: pin llama.cpp to the one-commit branch that forwards the…
registry: follow the llama.cpp pin (Fixes #1461)
package: the payload is checked for build-machine paths on every pack…
third_party: pin llama.cpp 6ea38d0 (main's pin + the path-map forward…
bong-water-water-bong commentedon Nov 8, 2026
Lemonade half measured after the merge of #1464: 126 → 0. Built
lemondfrom the merged script:CMAKE_CXX_FLAGSin the cache carries all four prefix maps (the evidence that mattered - cmake does not echo them, so the cache is the only place a silent-empty flag shows up), andstrings -a bin/lemond | grep -o '/home/' | wc -l= 0 against the 126 that shipped in c32 and c33; the binary is 18,278,448 B vs 18,290,736 B, 12,288 B smaller.\n\nThe Loom half remains (7 strings inlibloomc.so.0), and it is the one that needs the llama.cpp fork: itsGGML_HRX_CMAKE_ARGSlist never forwardsCMAKE_C_FLAGS/CMAKE_CXX_FLAGSinto the nestedggml-hrx-deps-build, so the engine's map stops at that boundary. So the payload is still not byte-reproducible and the qualification indocs/releases.mdstays until both land.bong-water-water-bong commentedon Nov 8, 2026
My comment above was measured at a superseded pin; the closure of this issue was right. I reported 'the Loom half remains' from
third_party/llama.cpp0ce01dafcf35. Main now pins6ea38d0457d3, which forwardsCMAKE_C_FLAGS/CMAKE_CXX_FLAGSinto the nested deps in the baseGGML_HRX_CMAKE_ARGS(lines 31-32), and main'scmake/hrx.cmakemaps${CMAKE_BINARY_DIR}=buildwhere Loom is fetched. So both halves are in the tree, and the Lemonade half is measured anyway:lemondrebuilt from the mergedbuild-lemonade.shcarries 0/home/occurrences against 126 shipped, and still names its sources relatively (117 lines beginninglemonade/src/, 119 containinglemonade/, 135 occurrences - counts of that build, in occurrences where the arrow is drawn).What an artifact cut before both halves still shows is the historical fact, not a defect: the c33 payload ships 146 build-machine paths (126
lemond+ 13lemonade+ 7libloomc). A roll from current main is expected to carry relative Loom paths and passbuild_path_strings; if that row reds, it is a finding about the fix rather than a fulfilled prediction.package: the payload ships the HSA runtime set it opens, and the rele…
pack-check: build_path_strings sees generated text, not only ELFs (th…
os: the BASE layer's group/other-writable modes are taken off, and a …