[laudney]

What this changes

A later --compact-at now replaces an earlier one. Before, a fraction after auto left automatic compaction on, so this one serve flag did not follow "the last value wins".

Command	main	This branch
--compact-at auto --compact-at 0.5	automatic compaction (the fraction is ignored)	fraction 0.5
--compact-at 0.5 --compact-at auto	automatic	automatic, and the stale fraction is cleared
one --compact-at auto or one --compact-at 0.25	as given	unchanged
--compact-at 1.5	refused	unchanged, same message

apply set compact_auto and never cleared it, and serve.zig prefers auto when both fields are set. Each occurrence now replaces both fields.

• Issue: Refs serve: read flags from a config file (--config FILE.toml), as vLLM and SGLang do #2882. This is the prerequisite named there: a config file's compact-at = "auto" under a command-line --compact-at 0.5 needs the last value to win. The defect is on the plain command line too, so it stands alone.
• Stacked on: main.
• atc overlap --json: run 2026-11-11 09:27 UTC on base 6b739dd. complete: false (the pull request and time limits of the snapshot). 4 conflicts (idol: Idol's denoiser on CUDA (stacked on #1766) #1767, ricursive: the publisher's DFlash2 drafter for Ripple-65B (stacked on #2231) #2260, ricursive: Ripple-65B-Preview on native Metal (NVFP4 and bf16 releases) #2231, ricursive: Ripple-65B-Preview on native CUDA (NVFP4) #2302), each in CHANGELOG.md only, where both sides add an Unreleased line; 4 clean; the rest untested by the snapshot. No source file conflicts.

Checks on this head

• Head commit: 907a9d6, rebased on main 7d97cc6 (was 9903293 on 6b739dd)
• Lean checker, python3 tools/zig/lean_check.py, on the base and on this head: 386 and 386, 0 new on the rebased head (tools/pr_check.py --base upstream/main: 0 errors). cli.zig goes from 587 to 589 lines.
• Tests run, with the command and pass, fail and skip counts:
  • zig build test-serve-flags-host (new filtered step): passes. The new test fails on main at expect(!fraction_last.compact_auto) and passes with the change.
  • zig build test-serve-flags-host test-server-cpu on this head: 1337 pass, 5 skip, 0 fail of 1342 (the server suite 1311 pass and 5 skip; the filtered step 21 pass).
  • zig fmt --check build.zig zig and git diff --check: clean.
• CI: to be read on this head after it runs.
• Not run, and why: zig build test test-golden in full and the macOS and Windows builds; the change is four lines in the flag parser, covered by the server host tests on Linux.

Receipt

n/a: the change is in the serve flag parser. It selects which compaction setting the server gets and reaches no kernel, token path or arithmetic.

Checklist

• The title and description match the head commit, including the receipt and the stack.
• New tests fail before the change, pass after it, and skip cleanly without their dependency.
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths in the diff or its commits. No AI attribution lines.

[ashhart]

Thanks @laudney, this is right. Each occurrence now replaces both fields, the test covers both orders plus the single values and the refusal, and its fraction-last case fails on main as it should. Ready from our side. It's queued behind the 1.2.7 release, which freezes main shortly, and we'll write the CHANGELOG line with its number and your thanks.

• Synthena (Ash's Agent)

[laudney]

Rebased on main 7d97cc656 as asked on #2885, so CI can run: head 907a9d61b. The only conflict was the Unreleased line in CHANGELOG.md; the code is unchanged. On this head, zig build test-serve-flags-host test-server-cpu gives 1337 pass, 5 skip, 0 fail, and lean is 386 on base and head. The description carries the new head and counts.

[ashhart]

Thanks @laudney for the rebase. 907a9d61b is green on CI and merges clean with today's main, so there's nothing more to do here. It's queued to land right after today's release, ahead of #2885.

• Synthena (Ash's Agent)

[ashhart]

@laudney, main had moved under this one, so we rebased it onto today's main ourselves and force-pushed (9f6fbc7a7c). The only conflict was the server's test-import list, where both sides are kept; its changelog line lands separately. test-serve-flags-host and test-server-cpu: 8/8 steps, 1,352/1,359 tests (7 skipped). It's queued to land once CI is green on the new head.

• Synthena (Ash's Agent)

[ashhart]

Thank you @laudney. This is merged into main with your commits under your own name, so you're listed as a TensorFold contributor. It ships in the next release, credited to you in the changelog and the release notes.

[ashhart]

Thanks @laudney! This is in TensorFold 1.2.8, out now: https://github.com/ashhart/TensorFold/releases/tag/v1.2.8

• Synthena (Ash's Agent)