[tawei-amd]

Fixes #900.

Problem

Gemma4 writes string arguments as raw text between <|"|> markers. The reader also accepts a plain " or ` as a closer (the model sometimes writes one instead of <|"|>), and took it as one whenever a , or a closer followed. Code like np.genfromtxt("train.csv", delimiter=",") or f("q","r") has that shape, so the value was cut short and the rest became a second, malformed call.

Fix

scan_gemma4_string (src/common/chat/chat_args.cpp): a plain quote now closes a string only if what follows is the next member (,key: or ,<|"|>), a closer, or the end, and the real <|"|> markers after it pair up (an odd one is the string's own closer). The real <|"|> keeps its old rule. One parser path serves every Gemma4-format model: gemma4e, gemma4-12b, gemma4e-flash, Nemotron-3.5-Bolt (same tool-call template as Google's).

Only the parser changes; no other component is touched.

Tests

• New src/test/chat_gemma4_args (no NPU, runs on Linux CI): the Gemma4 tool-call parsing splits a call when a string argument contains a double quote (v1.0.11) #900 shapes, plain-quote closers that must keep working, streaming prefixes, a seeded random sweep. Fails 10 checks on the old code, passes now. chat_continue still passes.

• Stress (local harness): random Gemma4-style calls with quote/comma/brace-heavy values. Old parser: 1353/2000 wrong. This PR: 0/200,000 (40 seeds), every streaming prefix and truncation checked; ASan/UBSan clean.

• Differential on benign calls incl. plain-quote closers and a missing opener: completed parses identical to the old code.

• Real models on the NPU (old portable build vs this PR, same 20 code-with-quotes prompts, temperature 0; a prompt passes if it gives exactly one well-formed file_write call):

  Model	old parser	this PR
  gemma4-it:e4b	8/20 (up to 6 calls per prompt)	20/20
  gemma4-it:e2b	8/20 (up to 4)	20/20
  gemma4-it:12b	8/20 (up to 4)	20/20
  gemma4e-flash:e4b*	8/20 (up to 6)	20/20
  gemma4e-flash:e2b*	8/20 (up to 4)	20/20
  Nemotron-3.5-Bolt-E4B	8/20	20/20

  * flash never calls a tool under tool_choice: auto (it answers in markdown), so these use tool_choice: required.
  Streamed results equal non-streamed on 10/10 (Bolt only).

• Perf: 560 KB worst case 61 ms (old 18 ms).

Trade-offs

• For a call that closes a string with a plain quote (malformed model output), streaming holds the arguments until the call ends; the final result is unchanged.
• Streaming was only checked on the NPU for Bolt; 12b was run non-streaming only.

🤖 Generated with Claude Code