split(#63)/2: stop sequences, thinking floor, role:tool 400

[antmikinka]

Summary

Stack 2/5 from mega PR #63 (after #64 merged into feat/openai-tools-server).

• OpenAI stop strings (block + stream) + unit tests
• Thinking budget floor 4096 + defaults + unit tests
• role:tool clear 400 for OWUI Memory/tools
• server.cpp single ownership unit for this stack slice

Base

Now feat/openai-tools-server (includes stack 1 / #64).

Stack

#	PR	Status
1	#64 ChatSession / EOS / pure	MERGED into tools
2	This PR	open
3	#66 load hygiene	open
4	#67 CI dual smoke	open
5	#68 docs	open

No MTP enable.

[antmikinka]

Retargeted base to feat/openai-tools-server now that #64 is merged. Branch includes merge of tools tip + server protocol commits. No self-merge.

[antmikinka]

Further comment-only cleanup on server.cpp (redundant stop/floor notes removed per explore QA list). Behavior unchanged. Cascaded to #66–#68 tips.

[antmikinka]

Same product belongs on main

Opened main retarget: #71

This stack PR still targets feat/openai-tools-server. Prefer landing #71 on main (human merge) once green.

CI fix pushed here too

thinking_budget_floor no longer rewrites explicit max_tokens (was 5→4096 → OOM kill on cpu/mac simple-math). Floor only applies when budget is unset.

[antmikinka]

Closing in favor of main retarget

This stacked PR (base: feat/openai-tools-server) is superseded for mainline by:

→ #71

Why close #65

• Product stop sequences / thinking floor / role:tool 400 should land on main, not only the tools long-lived branch (same pattern as split(#63)/1: ChatSession multi-turn, EOS merge, pure-graph opt-in #64 → fix(chat): multi-turn ChatSession re-prefill + EOS merge (from #64) #70).
• fix(server): stop sequences, thinking floor, role:tool 400 (from #65) #71 is the pure product slice of this PR on main (no generate.cpp / pure-graph risk; independent of fix(chat): multi-turn ChatSession re-prefill + EOS merge (from #64) #70 ChatSession).
• Thinking-floor policy fix (never overwrite explicit client max_tokens) is on fix(server): stop sequences, thinking floor, role:tool 400 (from #65) #71; that unblocks simple-math CI (was 5→4096 OOM on cpu/mac).

What was here

Area	Status on #71
stop sequences + tests	yes
thinking budget floor (explicit wins)	yes
role:tool 400	yes
ChatSession / EOS	not this PR — see #70
dual-lane CI	still stack #67 (retarget main separately if needed)

Stack note

Original stack after #64 on tools. Prefer human merge of #71 → main when green. Do not re-open this PR for mainline work; open new main retargets for later slices (#66 load hygiene, #67 dual-lane) if those should also go main.

Closing as superseded by #71.

[antmikinka]

Superseded by #71 (main retarget of this product slice). See closing comment for detail.