LiteRT-LM reason-faithfulness bypass driver + fix native SIGTRAP on long generations

[ber4444]

Summary

Adds a standalone :litert-eval Gradle module that drives the on-device LiteRT-LM (Qwen3-0.6B-int4) generator over the golden Move Coach cases and emits raw outputs as JSON, to be scored by ferryman's reason-faithfulness scorer. Generation and scoring are deliberately decoupled — two repos, two languages, one JSON contract.

Also fixes a native crash (JVM exit 133 / SIGTRAP) that prevented the driver from completing.

The crash this fixes

./gradlew :litert-eval:run --args="10 litert-outputs.json" was dying at case 8:

> Process '.../bin/java' finished with non-zero exit value 133
  (this value may indicate that the process was terminated with the SIGTRAP signal)

Exit 133 = 128 + 5 = SIGTRAP, a C++ assertion (CHECK failure) inside the litertlm-jvm native runtime — not a Kotlin exception, which is why try/catch (Throwable) in the driver couldn't intercept it and the whole JVM died.

Root cause: LitertLmTextGenerator requested maxNumTokens = 2048 (DEFAULT_CONTEXT_SIZE), but the Qwen3 model's KV cache tops out at 1280 (ekv1280 in the filename; max_num_tokens: 1280 in its metadata). Compiles fine, but the native runtime traps once a generation's prompt + Qwen3's <think> chain-of-thought + output crosses 1280 tokens — typically on checkmate cases. Case 8 is 1.f3 e5 2.g4 Qh4# (Fool's Mate), whose mate explanation triggers long reasoning.

Fix: lower DEFAULT_CONTEXT_SIZE from 2048 → 1024 (headroom under the 1280 cap). This also fixes the same latent crash in the production desktop Move Coach, which shares LitertLmTextGenerator.

Note for reviewers: the context-size change affects the live desktop Move Coach feature, not just the eval driver. The old value (2048) was unachievable against this model — the model can only honour 1280 — so it was a latent crash in the live feature too, not just the eval.

This is a different crash from the one already documented in evals/docs/litert-faithfulness-bypass.md (the NoSuchMethodError: SendChannel.close$default coroutines mismatch). That one is already fixed by the resolutionStrategy.force to coroutines 1.11.0 in litert-eval/build.gradle.kts — cases 1–7 were generating successfully, which proves the coroutines bridge works.

Why a separate :litert-eval module

litertlm-jvm 0.14.0 is internally inconsistent: its bytecode calls SendChannel.close$default (a static bridge that only exists in kotlinx-coroutines 1.11.0+) while its POM declares 1.9.0. :evals pulls Ktor via :server, which drags coroutines to 1.10.2 — no version in that graph provides the bridge. :litert-eval depends on :onDeviceAi + :coachapi only and forces coroutines to 1.11.0, the single version that satisfies litertlm's bytecode. Full rationale in evals/docs/litert-faithfulness-bypass.md.

Fixture

Normalize candidates.json tag vocabulary to the MoveCoachFallback constants (develops, center-control, …) so the prompt hands the model translated phrases, not raw strings (development, center) that miss describeTags' when branches.

Verification

./gradlew :litert-eval:run --args="10 litert-outputs.json"

	before	after
result	exit 133 (SIGTRAP) at case 8	EXIT=0, 10/10 cases
cases completed	7	10
route=litert records	0 (file not written)	10

Next step (ferryman repo): python3 eval_harness/score_litert_outputs.py /path/to/litert-outputs.json --write scorecard-litert.md

Changed files

• litert-eval/ — new module: build.gradle.kts (coroutines isolation + force), EvalLiteRtDriver.kt (the driver), .gitignore (excludes generated litert-outputs.json)
• onDeviceAi/.../LitertLmTextGenerator.kt — DEFAULT_CONTEXT_SIZE 2048 → 1024 (the crash fix)
• evals/docs/litert-faithfulness-bypass.md — design doc
• evals/golden/candidates.json — tag vocabulary normalization
• settings.gradle.kts — include :litert-eval