[Nir-Az]

Overview: Moves the system-nlohmann_json LibCI job from Ubuntu 24.04 to 26.04, so the unit tests run against a current toolchain instead of being untested there — and fixes the floating-point contraction that the move exposed.

Why

#15384 was a test-types-pose failure that only appears on Ubuntu 26.04, and it was invisible to this repo's CI. Only two Linux jobs run the untagged unit-test suite:

job	runner	filter	runs test-types-pose
U22_SH_Py_CI	ubuntu-22.04	--context "linux"	yes
U24_SH_Py_CI_SYS_JSON	ubuntu-24.04	--context "linux"	yes
U22_U24_SH_Py_DDS_CI	22.04, 24.04	--tag dds	no

U-Latest_ST_Py_EX_CfU does not help: it has no LibCI step, never passes -DBUILD_UNIT_TESTS (default OFF), and ubuntu-latest still resolves to ubuntu-24.04 on GitHub runners today.

The failure itself came from GCC 15.2 on 26.04 contracting multiply-adds into FMAs under the default -ffp-contract=fast, which shifted an intermediate rounding by one ULP and broke an exact float comparison in the test. That is the class of breakage this job is meant to catch — it surfaces on a toolchain bump, not on a code change.

Moving the job promptly surfaced a second instance of the same class, in post-processing/pytest-filters.py, which this PR also fixes.

Summary

• U24_SH_Py_CI_SYS_JSON becomes U26_SH_Py_CI_SYS_JSON, runs-on: ubuntu-26.04. All steps unchanged.
• Untagged Linux coverage becomes 22.04 + 26.04, bracketing 24.04.
• 24.04 keeps build coverage (U22_U24_ST_Py_EX_CfU) and DDS-tagged LibCI (U22_U24_SH_Py_DDS_CI). What it loses is the untagged suite.
• No new job, and no ruleset change — only LRS_libci_trigger is a required context.
• CMake/unix_config.cmake builds with -ffp-contract=off.

Chose the flip over adding a 26.04 matrix entry to keep the job count flat. The matrix alternative costs one extra job and would preserve untagged 24.04 coverage, if that tradeoff is preferred.

The -ffp-contract=off fix

pytest-filters.py compares post-processing output byte-for-byte against a reference dataset. The spatial and temporal filters both compute x * alpha + y * (1 - alpha) in float (src/proc/temporal-filter.h, src/proc/spatial-filter.{h,cpp} — nine sites). GCC and Clang default to -ffp-contract=fast, which fuses one of those products into an FMA wherever the ISA has one, dropping an intermediate rounding. The 1-ULP drift is written back into the temporal filter's history buffer and compounds across a frame sequence; a few pixels per frame survive truncation to uint16 as a ±1 depth difference.

This is not specific to Ubuntu 26.04. FMA is in the base aarch64 ISA, so the same four cases already differ on ubuntu-24.04-arm with GCC 13.3 — identical cases, identical pixel counts. Jetson and Apple-Silicon builds have been producing filter output 1 LSB away from x86 and Windows all along; it went unnoticed because the Jetson LibCI stage runs --live, which skips this test. The 26.04 job made an existing cross-platform inconsistency visible rather than creating one.

-ffp-contract=off restores bit-identical results across toolchains. Where the ISA has no FMA — x86 up to 24.04, and Windows/MSVC — there is nothing to fuse, so codegen is unchanged and existing byte-exact jobs are unaffected. A source-level fix was rejected: -ffp-contract=fast ignores statement boundaries, so splitting the expression into temporaries does not reliably prevent fusion, and =on still fuses because it is a single expression.

Only the unix_config.cmake platforms are changed, since those are the ones measured here.

Test plan

• Reproduced CI: Test failure when using the ubuntu 26.04 distribution #15384 on an ubuntu-26.04 runner and confirmed the merged test fix resolves it — three arms differing only in the line under test (upstream pose.h + exact compare: fail; contributor pose.h + exact compare: fail; + approximate compare: pass)
• Confirmed the mechanism on the runner: gcc 15.2.0, -ffp-contract=fast, __FP_FAST_FMA defined; measured error 4.77e-07 against an eq() tolerance of 2.03e-06
• Isolated the pytest-filters failure to FMA contraction with a six-arm experiment on ubuntu-26.04: -ffp-contract=off byte-exact; -fno-tree-vectorize and -O2 still failing (so not vectorization or optimization level); gcc-13 on 26.04 byte-exact; both runner CPU models present in passing and failing arms
• Confirmed the same four cases fail on ubuntu-24.04-arm (aarch64, GCC 13.3) with identical pixel counts, and go byte-exact with the flag
• Verified the flag reaches the real compile line from os_set_flags (not a command-line override) and that all fourteen cases are byte-exact on ubuntu-26.04, ubuntu-24.04-arm, and ubuntu-24.04
• This PR's own U26_SH_Py_CI_SYS_JSON job green — first full-suite run on GCC 15.2, so unrelated failures are possible
• LibCI run

---

🤖 Generated by AI

[sysrsbuild-gh-agentic]

🤖 Auto-generated by rs-agentic bot\n\nCI change looks correct — job rename and runner bump are consistent, no needs: references broken.\n\nTwo pre-existing patterns worth watching on 26.04:\n1. libglfw3 package (Prebuild step): transitional package renamed to libglfw3t64 since 24.04 — may be dropped on 26.04, breaking the apt step. Consider libglfw3t64 or confirm availability on the 26.04 runner.\n2. pip install without venv (Prebuild + LibCI pytest steps): PEP 668 externally-managed-environment may block bare pip calls on 26.04. Consider --break-system-packages or a venv.

[Nir-Az]

🤖 Auto-generated by rs-agentic bot\n\nCI change looks correct — job rename and runner bump are consistent, no needs: references broken.\n\nTwo pre-existing patterns worth watching on 26.04:\n1. libglfw3 package (Prebuild step): transitional package renamed to libglfw3t64 since 24.04 — may be dropped on 26.04, breaking the apt step. Consider libglfw3t64 or confirm availability on the 26.04 runner.\n2. pip install without venv (Prebuild + LibCI pytest steps): PEP 668 externally-managed-environment may block bare pip calls on 26.04. Consider --break-system-packages or a venv.

Wrong, it tun and pass