ds4-agent: fix every tool call failing on DeepSeek without --vision (+ regression test)

[rinaldofesta]

Since 4771329 ds4-agent cannot complete a single tool call on a DeepSeek text model. The first tool result dies with

COMPACTING tool result would exceed context: summarizing durable task state
ds4-agent: context full after compaction

at 1.9k of 262k context, on a 29-entry list .. Reproduced on an M5 Max 128 GB with DeepSeek V4 Flash 0731 IQ2_XXS, no --vision.

Root cause: agent_tool_observation_build sends every tool observation through ds4_chat_append_multimodal_message, whose guard rejects any model that is not GLM unless a DeepSeek vision encoder is loaded (fc8bf3c relaxed only that case). The observation never builds. agent_tool_observation_fits returns false on a build error as well, so the agent reads it as a context overflow, compacts, and gives up.

Fix, 8 lines in ds4.c: a user or tool message with zero images delegates to ds4_chat_append_message, the rendering the agent used before 4771329 (DeepSeek: user role plus <tool_result>; GLM: the same tokens the multimodal branch emits). The vision guard is untouched and still applies to messages carrying images.

Test, tests/ds4_test.c, --chat-multimodal-text-only: asserts token-identical output between the two entry points for the tool and user roles. On the engine before this patch it fails twice per role (assertion failed: ok == 1); with the patch it passes.

Validation on a MacBook Pro M5 Max 128 GB, Metal, Flash 0731 IQ2_XXS/Q2_K:

• ./ds4_test --chat-multimodal-text-only: before 4 assertion failures, after OK
• ./ds4_test --server OK, ./ds4_test --tool-call-quality OK, ./ds4_agent_test OK
• ds4-agent --non-interactive with one list tool call in a scratch project: no compaction, correct answer. An interactive session then ran 11 tool calls on a Next.js repo without a hitch.
• make ds4_cpu.o clean. The patch applies on current main b0a147a.

Not run: CUDA, ROCm, GLM models, --logprob-vectors.

Overlap with #924: same idea for the ds4.c hunk. #924 no longer applies to main after fc8bf3c, adds a second copy of the dispatch in ds4_agent.c, and has no test. If you would rather land #924 rebased, I can move the test there instead.

Two things I left out on purpose, both separate changes: agent_tool_observation_fits still reports any build error as a context overflow and drops the error text, and the image-bearing tool branch renders DeepSeek observations with GLM tags and no role token.

Implemented with the help of a coding agent (Claude Code) and reviewed locally by the author.

[rinaldofesta]

Unrelated follow-up in #965 (Metal DSpark scheduler bypass). The two do not touch the same lines and merge in either order.