fix: retain Android Dawn compatibility for LiteRT 0.17.1

[leehack]

LiteRT-LM v0.17.1 leaves the Android Dawn inputs unchanged, but the version-keyed override table currently drops the existing arm64/x64 correction when selecting that tag. This retains the exact checksummed overrides for v0.17.1 and extends compatibility/version-floor tests across the existing nine target configurations.

This is compatibility preparation, not qualified v0.17.1 adoption. No runtime/consumer/qualification pins change. The Windows DXIL fix remains separate. No issue closure is requested.

Validation: 183 local tooling tests pass with one platform-dependent skip; 21 diagnostics tests pass. Independent exact-head review at 2128df3 against 3eb4079 found no blocking code findings and independently passed 31 focused tests plus per-architecture checksum-rejection/preservation, provenance, and unknown-version checks. Exact-source component probes passed 13 Rust tests and 48 llguidance numeric grammar controls; production Rust/C++ conversion preserves integers above 2^53 and i64 limits.

Remaining qualification gates:

• Exact v0.17.1 tag 5e58e9a0aef7abf7091207a8b1d1063a1c800f08 has a syntax defect in llg_fc_tool_calls_test.cc:652 and mismatched new test delimiters. No repaired test is credited as an upstream pass. A separately reviewed test-only candidate removes the dangling expression and corrects eight delimiter triplets without changing assertions/numeric expectations; its full FC target passes22tests. The unchanged full target fails locally and in upstream CI.
• Official Apple hashes are identical to v0.17.0: iOS c94fc12aa0403cb47208e419cc3bfe258214ea17035f7a63c16de536869f2186; macOS 83efd536485c9d58fcd7fb7d4556ddb16ca46bb775b0449d08d9825c6836c1a4. They cannot establish inclusion of the integer fix without provenance or direct behavior.
• Python number grammar accepts exponent notation, while the unchanged Python parser rejects 1e3; this existing mismatch remains recorded.
• All 25 lifecycle tests were blocked locally by the Bash 4+ prerequisite (installed Bash 3.2). The unchanged gate now passes on the exact PR head in hosted Publication lifecycle: 25 tests, OK.
• Fresh macOS arm64 source-runtime build and exported-symbol checks now pass. Frozen public Dart CPU Gemma4 E2B probes pass exact integer tool calls for 1000, >2^53 and both i64 limits, plus cancellation/unload/reload/dispose. Qwen CPU text/lifecycle executes but produces no tool calls. Fresh ASR, other platforms/backends, incremental tool fragmentation and native leak instrumentation remain unqualified. Consumer evidence uses base6c01a69af plus a captured typed-tool-result diff, frozen131file hashes; not a pristine consumer commit claim.

The ordinary automatic PR workflow performs nine-target qualification using its unchanged v0.17.0 source pin. Those checks are baseline regression evidence, not v0.17.1 runtime qualification. No manual workflow dispatch, artifact publication, or downstream adoption is part of this PR.

Automatic hosted CI completed successfully on 2026-09-18 at exact head 2128df3c3da431f7daf66e2866aac4d7131d322f against base 3eb4079397d19e2058e176fcdb5ab15b28a70ad0: all 17 checks passed, including nine build/smoke targets, candidate verification, SentencePiece sanitizers, publication lifecycle, and Qwen inference on Linux x64, Windows x64, and macOS arm64. Qualification run. Final review-thread query returned no threads. PR #53 remains draft. These are v0.17.0 baseline checks; the v0.17.1 gates above remain open.

Follow-up qualification (separate from this preparation patch): a downloaded official v0.17.1 macOS archive matching SHA256 83efd536485c9d58fcd7fb7d4556ddb16ca46bb775b0449d08d9825c6836c1a4 fails a direct arm64 production-parser probe. All 18 integer cases across FC/Python/JSON serialize as doubles, and 9007199254740993 rounds down. Independent review verified exact archive bytes, library loading, and symbol binding. This confirms the macOS prebuilt lacks the promised integer behavior; it is component evidence, not public Dart/model or iOS qualification. PR marked ready only for its compatibility-preparation scope after fresh independent review; no adoption or merge claim.