fix(hot-reload): make it work on Linux, and pin down the one-flecs invariant by andrewgazelka · Pull Request #1079 · hyperion-mc/hyperion

andrewgazelka · 2026-07-29T14:33:18Z

Makes crates/hyperion-hot-reload work on Linux, and establishes the linking invariant the whole design rests on. Groundwork for hot reloading a game module on ix apply without disconnecting players; the deployment half is designed and written up but deliberately not in this PR.

It did not work on Linux, in two quiet ways

docs/hot-reload.md said "only tested on aarch64-darwin". That undersold it.

Exports. rustc links a dylib with its own anonymous version script ending in local: *, which demotes every symbol it did not generate. flecs's C symbols arrived with DEFAULT visibility and LOCAL binding — present and unreachable — so the demo host would not even link:

rust-lld: error: undefined symbol: ecs_progress
rust-lld: error: undefined symbol: ecs_ensure_id

--export-dynamic and --export-dynamic-symbol=ecs_* both leave the exported count at exactly 0; a version-script demotion is not something either can reverse. A second version script works, because ld merges them and an explicit pattern beats a * wildcard. Before: 9001 exported, 0 of them ecs_*. After: 10717 exported, 666 of them ecs_*, ecs_init GLOBAL.

One flecs, or the world is indexed two ways. flecs_ecs's derive emits a static INDEX per component type, initialised from a process-global pool, and that index is a slot in the world's component array. The flecs_manual_registration note in the doc is about the id being per-world; the index is not. Two copies of flecs_ecs is two pools — and nothing detects it. AbiToken passes, no error is raised, and both sides stay internally consistent while disagreeing about which slot is which.

What this PR contains

crates/hyperion-hot-reload/build.rs — the ELF version script (ENG-11272).
crates/hyperion-hot-reload/demo/index-probe-{host,module} — a probe for the shared-pool invariant.
Repin flecs_ecs onto andrewgazelka/Flecs-Rust@f09dc53, which gives that crate crate-type = ["dylib", "rlib"] and moves the flecs export script to where the dylib is actually produced.
docs/hot-reload.md — the findings, the traps, the build recipe, and the deploy design.

The probe is behavioural, because the two obvious tests both lie

Recorded in the code and the doc, because both look like success:

Comparing ecs_init as usize across the boundary reports a mismatch even when the copy is shared. An executable taking the address of a dynamically-linked function gets its own PLT stub. Measured shared and separate copies; both printed different addresses.
Comparing one type's index for equality passes when nothing is shared. Two independent pools each start at 1, so the first type registered on each side reads 1 and 1. This is exactly what the probe reported before the module referenced the runtime crate at all.

Allocation order cannot coincide. With one pool an index taken in the module is strictly greater than every index the host took first.

Measurements

hyperion as a plain rlib — separate pools. The host creates the second copy, pulling flecs_ecs in through hyperion's rlib while the module resolves it from the runtime dylib; a module that touches no hyperion component does not avoid this:

host indices: [1, 2, 3, 4] (max 4)
module's own type index: 1
hyperion::simulation::Position index: host 4, module 2
SHARED_POOL=false

With flecs_ecs and hyperion both dylibs, against the real pin with no local patch:

host indices: [1, 2, 3, 4] (max 4)
module's own type index: 5
hyperion::simulation::Position index: host 4, module 4
SHARED_POOL=true
SHARED_HYPERION_INDEX=true
PROBE_OK

demo.sh on x86_64-linux now matches darwin case for case: code-only change accepted with state intact, layout change refused with the world still ticking on the previous build, migration accepted rewriting 3 instances (21u32 becoming 22.0f32, converted rather than reinterpreted). Exit 0.

Negative control: reverting build.rs reproduces the link failure, so the fix is load-bearing rather than incidental.

Verified on dev-compute-6 (x86_64-linux, rustc 1.99.0-nightly dc3f85158) and re-verified on aarch64-darwin.

What is NOT in this PR, stated plainly

hyperion is not yet crate-type = ["dylib", "rlib"]. The measurement above required it, but it is not part of this change, because making it a dylib needs -C prefer-dynamic -C link-arg=-Wl,--undefined-version -C link-arg=-Wl,--allow-shlib-undefined across the build and that is a packaging decision, not a one-line edit. Recipe and rationale are in the doc.
No hyperion module is reloadable. No export_module!, no host/rules split of smash, no reload loop. app.run() is still flecs's own main loop with no per-tick hook.
The deployment half is designed, not shipped. reloadTriggers, the /etc indirection and the ExecReload client are written up in the doc and not built.
The probe fails on any build that does not use the dylib recipe. It is a binary, not a test, so nothing runs it in CI today. It should become a Linux check once hyperion is a dylib; adding it before then would just be red.
The --allow-shlib-undefined requirement has a tidier fix I did not take. simulation/metadata/mod.rs:212 hand-writes impl PartialOrd for $name where $type: PartialOrd, unsatisfiable for 7 metadata types because glam's Quat and Vec3 have no PartialOrd. rustc never codegens those partial_cmp bodies but still exports them. They can never be called, so allowing them undefined is sound — but removing the blanket impl would remove the need for the flag.

Related: ENG-11272 (this), ENG-11279 (devShell cannot build smash on Linux; jemalloc 5.3.1 vs GCC 15 — pre-existing, unrelated, does not affect the nix build path).

🤖 Generated with Claude Code

--export-dynamic cannot undo what rustc does to a dylib. rustc links one with its own anonymous version script ending in local: *, which demotes every symbol it did not generate, so flecs C symbols sit in the object with DEFAULT visibility and LOCAL binding -- present and unreachable. Measured on x86_64-linux: 9001 exported symbols, zero of them ecs_*, ecs_init reading FUNC LOCAL DEFAULT. --export-dynamic-symbol=ecs_* is equally powerless against a version-script demotion; both were tried. The crate was therefore not merely untested on Linux, it did not work there at all. The demo host fails to link: rust-lld: error: undefined symbol: ecs_progress rust-lld: error: undefined symbol: ecs_ensure_id ld merges multiple version scripts and an explicit pattern beats a * wildcard, so a second script naming these globs promotes exactly them and leaves rustcs own exports alone. After: 10717 exported, 666 of them ecs_*, ecs_init GLOBAL, and demo.sh runs all three cases on x86_64-linux with the same output as aarch64-darwin -- code-only change accepted with state intact, layout change refused with the world still ticking on the old build, migration accepted rewriting 3 instances. Darwin re-verified unchanged.

…dylib boundary Everything the reload gate does assumes the host and a module dylib draw component indices from one pool. Nothing in the reload path checks it, and both sides are internally consistent when they do not, so the world is simply indexed two different ways and no error is raised. The probe is behavioural rather than an address comparison. Comparing ecs_init as usize across the boundary reports a mismatch even when the copy is shared, because an executable taking the address of a dynamically linked function gets its own PLT stub; measured that trap directly, with separate and shared copies both printing different addresses. Allocation order cannot be faked: with one pool, an index taken in the module is strictly greater than every index the host took first. It also catches a false positive worth recording. Before the module referenced the runtime crate at all it linked its own static copy of everything, and two independent pools each starting at 1 made hyperion::simulation::Position read index 1 on both sides. That looks exactly like success. The measurement that distinguishes them is allocation order, not equality. Result with hyperion as a plain rlib -- separate pools: host indices: [1, 2, 3, 4] (max 4) module own type index: 1 Position index: host 4, module 2 SHARED_POOL=false The host is what creates the second copy: it drags flecs_ecs in through hyperions rlib while the module resolves it from the runtime dylib. Result with flecs_ecs and hyperion both dylibs -- one pool: host indices: [1, 2, 3, 4] (max 4) module own type index: 5 Position index: host 4, module 4 SHARED_POOL=true SHARED_HYPERION_INDEX=true PROBE_OK Verified on x86_64-linux (dev-compute-6).

…eploy shape The document claimed the crate was merely untested on Linux. It did not work there, in two independent ways, and both are quiet: exports demoted to LOCAL by rustcs own version script, and a second flecs copy that leaves host and module indexing one world two different ways with no error from anything. Records the two measurements that give the wrong answer (comparing ecs_init addresses across a PLT boundary, and comparing one types index for equality when two pools both start at 1), because both look like success. Also writes down the deploy mechanism: NixOS reloads rather than restarts a unit when only X-Reload-Triggers changed, which is the whole basis for a game-logic change reaching a running server without disturbing a player. States plainly that it is not built.

andrewgazelka/Flecs-Rust f09dc53 gives flecs_ecs crate-type ["dylib", "rlib"] and a build script re-exporting flecs C symbols from the dylib. Hot reloading needs exactly one copy of that crate in the process, because it owns the process-global pool handing out each component types index into a worlds component array. Verified against the pin with no local patch, on x86_64-linux: host took component indices 1..4, the module then took 5, and hyperion::simulation::Position read index 4 on both sides. PROBE_OK. Darwin re-verified: demo.sh exits 0 with the same three verdicts.

github-actions · 2026-07-29T14:34:25Z

Benchmark Results for general

ray_intersection/aabb_size_0.1                     [  18.6 ns ...  18.7 ns ]      +0.21%
ray_intersection/aabb_size_1                       [  18.8 ns ...  18.7 ns ]      -0.21%
ray_intersection/aabb_size_10                      [  18.6 ns ...  18.7 ns ]      +0.09%
ray_intersection/ray_distance_1                    [   1.3 ns ...   1.3 ns ]      -0.60%
ray_intersection/ray_distance_5                    [   1.3 ns ...   1.3 ns ]      +0.12%
ray_intersection/ray_distance_20                   [   1.3 ns ...   1.3 ns ]      +0.24%
overlap/no_overlap                                 [  15.4 ns ...  15.5 ns ]      +0.17%
overlap/partial_overlap                            [  15.5 ns ...  15.5 ns ]      +0.16%
overlap/full_containment                           [  14.5 ns ...  14.5 ns ]      -0.33%
point_containment/inside                           [   5.4 ns ...   5.4 ns ]      -0.06%
point_containment/outside                          [   5.7 ns ...   5.7 ns ]      -0.02%
point_containment/boundary                         [   5.4 ns ...   5.4 ns ]      -0.00%

Comparing to dd8311f

Allows print_stdout in the probe host, matching the demo host: a binary whose whole purpose is reporting a measurement. Drops the hyperion-hot-reload dependency from both probe crates. It was unused by cargo-machete's reckoning and, more to the point, the comment claiming that reference was what made the two sides share one flecs was wrong once flecs_ecs became a dylib. Removing it entirely leaves the probe passing: host indices: [1, 2, 3, 4] (max 4) module own type index: 5 Position index: host 4, module 4 PROBE_OK The doc comment now says what actually makes the pool shared, and records that a single index reading equal on both sides is not evidence: two separate pools both start at 1.

github-actions · 2026-07-29T14:39:29Z

Benchmark Results for general

ray_intersection/aabb_size_0.1                     [  18.6 ns ...  18.6 ns ]      +0.02%
ray_intersection/aabb_size_1                       [  18.6 ns ...  18.7 ns ]      +0.19%
ray_intersection/aabb_size_10                      [  18.6 ns ...  18.7 ns ]      +0.24%
ray_intersection/ray_distance_1                    [   1.3 ns ...   1.3 ns ]      -0.14%
ray_intersection/ray_distance_5                    [   1.3 ns ...   1.3 ns ]      +0.22%
ray_intersection/ray_distance_20                   [   1.3 ns ...   1.3 ns ]      +0.20%
overlap/no_overlap                                 [  15.9 ns ...  15.9 ns ]      +0.04%
overlap/partial_overlap                            [  15.9 ns ...  16.0 ns ]      +0.71%*
overlap/full_containment                           [  14.5 ns ...  14.9 ns ]      +2.79%*
point_containment/inside                           [   5.4 ns ...   5.4 ns ]      -0.17%
point_containment/outside                          [   5.7 ns ...   5.7 ns ]      +0.25%
point_containment/boundary                         [   5.4 ns ...   5.4 ns ]      +0.13%

Comparing to dd8311f

The repin commit left `crate-type = ["dylib", "rlib"]` on `hyperion`. That is required for the shared component-index pool, but only together with the rest of the recipe -- `-C prefer-dynamic`, `-Wl,--undefined-version`, `-Wl,--allow-shlib-undefined` -- which a plain `cargo test` does not use. Without them the dylib fails to link: ld: Undefined symbols for architecture arm64 ld: symbol(s) not found for architecture arm64 so the whole workspace stopped building. `cargo test --workspace` now passes again. Making `hyperion` a dylib is a packaging decision affecting every consumer, and it does not belong in a fix PR. The recipe and the measurements that justify it are in docs/hot-reload.md. The probe consequently cannot pass on a default build, which is the honest state rather than a regression: it exists to detect exactly the configuration the repo currently has. Its failure message now says so and points at the recipe instead of just reporting two numbers.

github-actions · 2026-07-29T14:44:48Z

Benchmark Results for general

ray_intersection/aabb_size_0.1                     [  18.7 ns ...  18.7 ns ]      -0.21%
ray_intersection/aabb_size_1                       [  18.8 ns ...  18.8 ns ]      +0.16%
ray_intersection/aabb_size_10                      [  18.8 ns ...  18.7 ns ]      -0.23%
ray_intersection/ray_distance_1                    [   1.3 ns ...   1.3 ns ]      -0.21%
ray_intersection/ray_distance_5                    [   1.3 ns ...   1.3 ns ]      +0.13%
ray_intersection/ray_distance_20                   [   1.3 ns ...   1.3 ns ]      -0.34%
overlap/no_overlap                                 [  15.6 ns ...  15.6 ns ]      -0.09%
overlap/partial_overlap                            [  15.9 ns ...  15.6 ns ]      -1.82%
overlap/full_containment                           [  14.6 ns ...  14.6 ns ]      +0.00%
point_containment/inside                           [   5.4 ns ...   5.4 ns ]      -0.03%
point_containment/outside                          [   5.7 ns ...   5.7 ns ]      +0.29%
point_containment/boundary                         [   5.4 ns ...   5.4 ns ]      +0.06%

Comparing to dd8311f

A rules-only change is roughly 2 s of compile and tens of milliseconds of reload; touching the engine is 4.42 s of compile and costs the process, the world and every connected player. That contrast is the argument for the whole design, and it was not written down anywhere. States what the numbers are not: debug rather than release, a whole binary relink rather than a rules dylib, and a 165 KB probe module rather than smash.

github-actions · 2026-07-29T14:48:20Z

Benchmark Results for general

ray_intersection/aabb_size_0.1                     [  22.0 ns ...  22.0 ns ]      +0.06%
ray_intersection/aabb_size_1                       [  22.0 ns ...  22.0 ns ]      +0.11%
ray_intersection/aabb_size_10                      [  22.0 ns ...  22.0 ns ]      +0.14%
ray_intersection/ray_distance_1                    [   0.9 ns ...   0.9 ns ]      -0.05%
ray_intersection/ray_distance_5                    [   0.9 ns ...   0.9 ns ]      +0.35%
ray_intersection/ray_distance_20                   [   0.9 ns ...   0.9 ns ]      -0.27%
overlap/no_overlap                                 [  14.6 ns ...  14.6 ns ]      -0.05%
overlap/partial_overlap                            [  15.0 ns ...  15.1 ns ]      +0.36%
overlap/full_containment                           [  14.0 ns ...  14.1 ns ]      +0.20%
point_containment/inside                           [   5.9 ns ...   5.9 ns ]      +0.37%
point_containment/outside                          [   5.9 ns ...   5.9 ns ]      +0.13%
point_containment/boundary                         [   5.9 ns ...   6.0 ns ]      +0.59%

Comparing to dd8311f

… list

github-actions · 2026-07-29T14:49:11Z

Benchmark Results for general

ray_intersection/aabb_size_0.1                     [  18.6 ns ...  18.7 ns ]      +0.31%
ray_intersection/aabb_size_1                       [  18.6 ns ...  18.6 ns ]      -0.17%
ray_intersection/aabb_size_10                      [  18.7 ns ...  18.6 ns ]      -0.18%
ray_intersection/ray_distance_1                    [   1.3 ns ...   1.3 ns ]      +0.17%
ray_intersection/ray_distance_5                    [   1.3 ns ...   1.3 ns ]      +0.38%
ray_intersection/ray_distance_20                   [   1.3 ns ...   1.3 ns ]      -0.16%
overlap/no_overlap                                 [  15.9 ns ...  15.8 ns ]      -0.24%
overlap/partial_overlap                            [  15.5 ns ...  15.5 ns ]      +0.16%
overlap/full_containment                           [  14.6 ns ...  14.6 ns ]      -0.05%
point_containment/inside                           [   5.5 ns ...   5.5 ns ]      -0.15%
point_containment/outside                          [   5.7 ns ...   5.7 ns ]      -0.10%
point_containment/boundary                         [   5.4 ns ...   5.4 ns ]      +0.35%

Comparing to dd8311f

codecov · 2026-07-29T14:58:31Z

Codecov Report

❌ Patch coverage is 0% with 41 lines in your changes missing coverage. Please review.
✅ Project coverage is 54.59%. Comparing base (dd8311f) to head (df28ffa).

Files with missing lines	Patch %	Lines
...erion-hot-reload/demo/index-probe-host/src/main.rs	0.00%	35 Missing ⚠️
...rion-hot-reload/demo/index-probe-module/src/lib.rs	0.00%	6 Missing ⚠️

@@            Coverage Diff             @@
##             main    #1079      +/-   ##
==========================================
- Coverage   54.65%   54.59%   -0.07%     
==========================================
  Files         361      363       +2     
  Lines       33257    33298      +41     
  Branches     1259     1259              
==========================================
  Hits        18178    18178              
- Misses      14793    14834      +41     
  Partials      286      286

Files with missing lines	Coverage Δ
...rion-hot-reload/demo/index-probe-module/src/lib.rs	`0.00% <0.00%> (ø)`
...erion-hot-reload/demo/index-probe-host/src/main.rs	`0.00% <0.00%> (ø)`

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

The repin changed the Cargo.lock source string, and outputHashes is keyed by that exact string, so every nix check failed to evaluate: error: outputHashes is missing hashes for git source strings in Cargo.lock: git+https://github.com/andrewgazelka/Flecs-Rust?rev=f09dc53... sha256-DlMOSY7NyoPoR8w4yswm3O97BegcNWcLl/fW3wOAmRs= from nix-prefetch-git --fetch-submodules at that rev.

github-actions · 2026-07-29T15:01:13Z

Benchmark Results for general

ray_intersection/aabb_size_0.1                     [  18.7 ns ...  18.7 ns ]      +0.22%
ray_intersection/aabb_size_1                       [  18.7 ns ...  18.8 ns ]      +0.13%
ray_intersection/aabb_size_10                      [  18.7 ns ...  18.7 ns ]      +0.05%
ray_intersection/ray_distance_1                    [   1.3 ns ...   1.3 ns ]      -0.17%
ray_intersection/ray_distance_5                    [   1.3 ns ...   1.3 ns ]      +0.29%
ray_intersection/ray_distance_20                   [   1.3 ns ...   1.3 ns ]      -0.24%
overlap/no_overlap                                 [  15.6 ns ...  15.6 ns ]      -0.39%
overlap/partial_overlap                            [  16.0 ns ...  16.0 ns ]      -0.12%
overlap/full_containment                           [  14.5 ns ...  14.5 ns ]      -0.02%
point_containment/inside                           [   5.4 ns ...   5.4 ns ]      -0.16%
point_containment/outside                          [   5.7 ns ...   5.7 ns ]      -0.02%
point_containment/boundary                         [   5.5 ns ...   5.5 ns ]      -0.11%

Comparing to dd8311f

…iour-only argument Three gaps a stranger would have hit. A behaviour-only module does not avoid the shared-pool requirement. The registration/behaviour split in CLAUDE.md makes it look like it should: a library that registers nothing cannot collide with anything. But registering a component and looking one up are different operations and only the first is avoided -- a system's query still resolves T through T::index(). The probe module registers nothing at all and still read Position as index 2 where the host had it at 4. The split is worth keeping for the hazard it does address, which is component layout, and that is now said explicitly along with the boundary it implies: changing what a system does is a reload, changing a component type is a rebuild and a restart. What makes the pool shared is the dependency being a dylib, not a consumer referencing it. An earlier probe carried a call to AbiToken::current() with a comment claiming otherwise; removing the dependency entirely leaves the probe passing. And the four remaining steps are now written down in order, with the packaging step named as the risky one and the reason why: every measurement here came from a cargo build, not a nix one. Also records that adoption needs no scheduled restart -- the fleet already restarts for version bumps, so the split host takes effect on the next apply happening for its own reasons -- and that the build tree on dev-compute-6 was deleted when the node was released.

github-actions · 2026-07-29T15:03:08Z

Benchmark Results for general

ray_intersection/aabb_size_0.1                     [  17.2 ns ...  17.2 ns ]      +0.26%
ray_intersection/aabb_size_1                       [  17.6 ns ...  17.5 ns ]      -0.52%
ray_intersection/aabb_size_10                      [  17.3 ns ...  17.2 ns ]      -0.19%
ray_intersection/ray_distance_1                    [   1.5 ns ...   1.5 ns ]      -0.34%
ray_intersection/ray_distance_5                    [   1.5 ns ...   1.5 ns ]      +0.30%
ray_intersection/ray_distance_20                   [   1.5 ns ...   1.5 ns ]      +0.04%
overlap/no_overlap                                 [  15.8 ns ...  15.8 ns ]      -0.00%
overlap/partial_overlap                            [  15.9 ns ...  15.9 ns ]      +0.09%
overlap/full_containment                           [  14.8 ns ...  14.9 ns ]      +0.17%
point_containment/inside                           [   6.0 ns ...   6.0 ns ]      -0.24%
point_containment/outside                          [   6.0 ns ...   6.0 ns ]      +0.13%
point_containment/boundary                         [   6.0 ns ...   6.1 ns ]      +0.19%

Comparing to dd8311f

Making flecs_ecs a dylib means it emits two artifacts on every build. I tried to measure whether that costs build time and could not confirm the comparison: 8875 ms with both against 8817 ms with the rlib alone is within noise, but a leftover dylib in the target directory means the rlib-only configuration may never have taken effect. Recording it as unmeasured rather than as zero, because the difference between those two claims is what decides whether the next person needs to look.

github-actions · 2026-07-29T15:19:45Z

Benchmark Results for general

ray_intersection/aabb_size_0.1                     [  18.7 ns ...  18.8 ns ]      +0.07%
ray_intersection/aabb_size_1                       [  18.7 ns ...  18.7 ns ]      -0.05%
ray_intersection/aabb_size_10                      [  18.7 ns ...  18.7 ns ]      +0.02%
ray_intersection/ray_distance_1                    [   1.3 ns ...   1.3 ns ]      +0.02%
ray_intersection/ray_distance_5                    [   1.3 ns ...   1.3 ns ]      +0.09%
ray_intersection/ray_distance_20                   [   1.3 ns ...   1.3 ns ]      -0.08%
overlap/no_overlap                                 [  16.5 ns ...  16.4 ns ]      -0.06%
overlap/partial_overlap                            [  16.1 ns ...  16.1 ns ]      -0.10%
overlap/full_containment                           [  14.7 ns ...  14.7 ns ]      -0.05%
point_containment/inside                           [   5.4 ns ...   5.4 ns ]      +0.43%
point_containment/outside                          [   5.7 ns ...   5.7 ns ]      -0.25%
point_containment/boundary                         [   5.4 ns ...   5.4 ns ]      -0.32%

Comparing to dd8311f

andrewgazelka added 4 commits July 29, 2026 07:08

andrewgazelka changed the title ~~hot-reload: make it work on Linux, and pin down the one-flecs invariant~~ fix(hot-reload): make it work on Linux, and pin down the one-flecs invariant Jul 29, 2026

github-actions Bot added the fix label Jul 29, 2026

docs(hot-reload): explain why two build scripts carry the same export…

738e0a9

… list

andrewgazelka merged commit 312bed6 into main Jul 30, 2026
10 of 11 checks passed

andrewgazelka deleted the hotreload branch July 30, 2026 03:19

Navigation Menu

Uh oh!

Uh oh!

fix(hot-reload): make it work on Linux, and pin down the one-flecs invariant - #1079

fix(hot-reload): make it work on Linux, and pin down the one-flecs invariant#1079
andrewgazelka merged 11 commits into
mainfrom
hotreload

andrewgazelka commented Jul 29, 2026

Uh oh!

github-actions Bot commented Jul 29, 2026

Uh oh!

github-actions Bot commented Jul 29, 2026

Uh oh!

github-actions Bot commented Jul 29, 2026

Uh oh!

github-actions Bot commented Jul 29, 2026

Uh oh!

github-actions Bot commented Jul 29, 2026

Uh oh!

codecov Bot commented Jul 29, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Jul 29, 2026

Uh oh!

github-actions Bot commented Jul 29, 2026

Uh oh!

github-actions Bot commented Jul 29, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Uh oh!

Conversation

andrewgazelka commented Jul 29, 2026

It did not work on Linux, in two quiet ways

What this PR contains

The probe is behavioural, because the two obvious tests both lie

Measurements

What is NOT in this PR, stated plainly

Uh oh!

github-actions Bot commented Jul 29, 2026

Benchmark Results for general

Uh oh!

github-actions Bot commented Jul 29, 2026

Benchmark Results for general

Uh oh!

github-actions Bot commented Jul 29, 2026

Benchmark Results for general

Uh oh!

github-actions Bot commented Jul 29, 2026

Benchmark Results for general

Uh oh!

github-actions Bot commented Jul 29, 2026

Benchmark Results for general

Uh oh!

codecov Bot commented Jul 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

github-actions Bot commented Jul 29, 2026

Benchmark Results for general

Uh oh!

github-actions Bot commented Jul 29, 2026

Benchmark Results for general

Uh oh!

github-actions Bot commented Jul 29, 2026

Benchmark Results for general

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

codecov Bot commented Jul 29, 2026 •

edited

Loading