WebGPU modelLoadFromUrl is a 408-line function whose retry ladder restarts itself by mutating a loop counter

[leehack]

Found during a code-health audit of main, confirmed by a follow-up verification pass. No fix applied.

Issue #354 lists webgpu_backend.dart in its file-size table, but its problem statement and all three suggested steps are about LlamaCppService and native-library discovery. This function is never described there, so filing separately.

What

WebGpuLlamaBackend.modelLoadFromUrl — lib/src/backends/webgpu/webgpu_backend.dart:971-1378, 408 lines.

Before the loop it validates params, resolves preferMemory64, resolves thread counts, applies a Safari GPU-layer clamp via _resolveSafeRequestedGpuLayers, and marshals the progressCallback across the JS boundary (:976, :981, :986-988, :991-1003, :1019-1038). Then it runs an attempt ladder from _buildLoadAttempts (:1041, loop at :1056) whose catch block escalates.

The part that makes it hard to change

Six retry branches, each ending in continue; — :1240, :1261, :1277, :1302, :1321, :1350 — behind five named predicates (shouldRetryWithSmallerRemoteFetchChunks, shouldRetryWithoutRemoteFetchBackend, shouldRetryWithWasm32, shouldRetryWithWasm64, canRetry) plus one inline fsWriteFailed && coreVariant == 'wasm64'.

Five of the six restart the entire ladder by assigning index = -1 — :1232, :1251, :1271, :1288, :1314 — mutating a for loop's induction variable from inside a catch. The escalation state is carried in seven mutable locals declared at :1048-1054 (retriedWithWasm32, retriedWithWasm64, retriedWithoutRemoteFetchBackend, remoteFetchChunkRetryCount, retriedAfterFsWriteFailureWithRemote, remoteFetchBackendKnownUnstable, wasm64InteropKnownBroken) plus two instance fields (_preferMemory64Override, _forceRemoteFetchBackendOverride) mutated at :981, :1030 and inside four branches.

To be clear about what this is not: it terminates. Every restart is latched by its own boolean and chunk-halving is capped at 10. This is a testability and maintainability problem, not a hang risk.

Half the ladder is untested

Branch	Coverage
wasm64 memory-pressure retry	Covered — webgpu_engine_multimodal_browser_integration_test.dart:317 asserts modelLoadCallCount == 2
wasm64 FS-write, both arms	Covered — same file, :334, :365
shouldRetryWithSmallerRemoteFetchChunks	No test
shouldRetryWithoutRemoteFetchBackend	No test
shouldRetryWithWasm32 (BigInt interop)	No test
canRetry ladder-advance	No test

Repo-wide greps for bigint, model_fetch_backend_abort, model_fetch_backend_skipped_small and thread_constructor_failed across test/ return zero hits. The covered branches do run in CI (.github/workflows/ci.yml:218, dart test -p chrome --exclude-tags local-only; the file is @TestOn('browser')).

dart analyze reports no issues here, so nothing guards against it growing further.

A duplication to fix while you are in there

The index-to-thread-count switch at :1067-1074 is copy-pasted verbatim into a log-string interpolation at :1343-1348, purely to print the number it already computed.

Suggested direction

_buildLoadAttempts (:632-671) already models the attempt sequence as data, so the usual "make it a table" advice is half-done. The part that is not data-driven is the failure-classifier and escalation layer above it.

1. Extract the pre-loop setup (validation, thread resolution, Safari clamp, callback marshalling) into its own function so the ladder starts at the top of the body.
2. Replace index = -1 with an explicit restart signal returned by a classifier — e.g. a function mapping (error, state) -> RetryDecision{advance | restartWith(overrides) | giveUp} — so the escalation policy becomes unit-testable without a browser.
3. Add coverage for the four untested branches once they are reachable from a test seam.
4. Collapse the duplicated thread-count switch.

Note on the audit's numbers

The original audit said "seven-branch". Six is the defensible count (six continue; statements); seven is reachable only by also counting the terminal _normalizeBridgeRuntimeError/rethrow path or the non-retrying state mutation at :1185. It also said 410 lines; measured span is 408.