Decode-quality primitives: scaffolding-token mask and argmax-is-EOG stop signal

[FuJacob]

Summary

Two additions for autocomplete decode quality, both zero cost on the hot sampling path and both source-compatible with the app's existing call sites (SamplingConfig untouched; SampleResult gains an appended field that callers only read).

Scaffolding-token mask (load time). buildTokenMasks now probes each token's special-rendered piece and adds single-token chat/instruct/FIM scaffolding (<|im_end|>, <start_of_turn>, [INST], FIM marker families) to the existing -inf logit-bias table when the GGUF ships them without the control attribute, plus an unflagged BOS. EOG tokens stay exempt so natural stops keep firing. Well-formed GGUFs flag these as control already, so the expected count is 0 on the catalog models; this is insurance against vocabularies that do not, surfaced via getMaskedScaffoldingTokenCount for tests and diagnostics.

argmax_is_eog on SampleResult. True when the raw distribution's single most-likely token at this position is an end-of-generation token. Stochastic sampling can draw past the point where the model wants to stop; this lets the caller detect the stop intent on the exact step it appears and finalize cleanly. Computed in C++ while the logits row is hot: one O(vocab) pass per sampled token (tens of microseconds; the row is unmutated because llama_sampler_sample works on a copied candidate array). The seed token's verdict is captured at decodePrompt while its logits row is still resident, mirroring seed_logprob.

Validation

swift build                       # clean
COTABBY_TEST_MODEL_PATH=<SmolLM2-135M> swift test
# 25 tests, 0 failures (includes 3 new: masked-count default,
# scaffolding-marker pieces never sampled at temp 1.8,
# argmax_is_eog == isEndOfGenerationToken(token) under greedy)
COTABBY_TEST_MODEL_PATH=<Qwen3.5-0.8B-Base> swift test --filter 'Argmax|Scaffolding'
# both new model-bound tests pass on a catalog base model

Note: testEndToEndWithModel has a pre-existing, model-specific failure on Qwen3.5-0.8B-Base (partial trimKV returns false on that model's attention layout; KV count assertion follows). Verified identical on unmodified main; unrelated to this change and the app already falls back to a fresh prompt build when trimKV reports failure.

Risk / rollout

Mask additions only ever remove tokens that should never appear as autocomplete text; EOG exemption is covered by tests. The argmax flag is informational until a consumer opts in app-side.