[joelteply]

The reasoning-budget-0 fix (#95, reviewed by Fable) on the integration line continuum pins. #95 merged into fork master, which is 99 commits behind continuum's pin 0d7dda6; every fork PR since #84 targets this branch. Same three commits, cherry-picked clean onto a8c33da:

• sampling: in FORCING, force_pos advances only on the forced token, and a natural end tag ends the forcing. A prefilled <think>\n can no longer consume the forced end, which had let the model think to its cap at budget 0.
• gguf.cpp: #include <cerrno>.
• Missing libc includes: the includes upstream ggml-org has and this line lacks (the WASM WebGPU build stopped providing them transitively).

The fork's default branch is now this line, so new PRs target what continuum ships.

Measured

Integration build 45ae996 (this head), IntelMac CPU, Qwen3.5-0.8B Q4_K_M, --jinja --reasoning-format deepseek -t 2, 2026-11-07 ~01:20Z. The sampler and its test are byte-identical to master's, but server-context.cpp differs by ~880 lines, so the probe was re-run on this build.

• test-reasoning-budget: OK (13 tests). Mutation (on master's build): reverting the fix aborts the suite at activate immediately budget=0.

request	deployed 0d7dda6	this head
schema, budget 0	thinks to the cap, content empty	reasoning empty, 78-token schema JSON, stop
no schema, budget 0	thinks to the cap	reasoning empty, answers from token 1
schema, budget 1 / 40 / none	bounded / forced end + JSON / thinks to the cap	unchanged
budget 0 + enable_thinking=false (continuum ggml-org#5808's pair)	reasoning empty (the template closes it)	reasoning empty, 95 tokens, stop

🤖 Generated with Claude Code

https://claude.ai/code/session_01FRQWzgSfo79JtnwZywHKrE

[joelteply]

PASS at 45ae996.

Verified: I diffed #95's squashed patch (25c5ec17b~1..25c5ec17b) against this PR's total patch (889580fc/seq-state-across-stream-counts..45ae99681). Every added and removed line is identical. The only differences are blob index hashes and one unchanged context line (#include <array> vs <memory> near an insertion), because the integration line's files differ around the edits.

Logic: the same FORCING-accept fix I passed on #95. It is re-measured on a server built from this head, because server-context.cpp differs by ~880 lines from master: unit 13/13; budget 0 gives empty reasoning and 78-token schema JSON with finish=stop; budgets 1, 40 and none are unchanged; and ggml-org#5808's two-lever pair also gives no think. Merge on green CI. Then bump the pin (f0a909bd), and set the fork's default branch to the integration line.

[joelteply]

PASS carries to 9bd4d63. The delta from 45ae996 is .github/workflows/build-cpu.yml +3 lines: a comment, plus if: github.repository == 'ggml-org/llama.cpp' on build-cmake-pkg, whose runner is self-hosted and exists only upstream. It fixes the root cause: every fork PR has waited 24 h on runner-less jobs that were then cancelled. Disabling the four self-hosted-only workflows in repo settings (reversible) covers the rest. No logic change.