fix(extension): unblock local proving via sdk re-entrancy + sync-mgr classification

[WiktorStarczewski]

Web SDK PR: #152

Summary

Three independent wallet-side fixes that, together with the linked web-sdk#152, unblock local proving on Chrome MV3 and stop the misleading "cannot reach the miden node" banner from firing during honest long proves.

1. SDK fix consumption — bumps @miden-sdk/miden-sdk to 0.14.10 via the linked-PR pipeline. The SDK fix is the load-bearing change: _withInnerWebClient was deadlocking the moment its callback called back into a proxied wasm method on the same client (inner.getInputNote(...), inner.executeTransaction(...), etc.). All wallet usage of proveLocallyViaOffscreen for consumes and sends hung indefinitely. With the SDK fix, the offscreen-doc MT-WASM prove path actually completes.
2. SyncManager classification (src/lib/miden/back/sync-manager.ts) — the 5 s SYNC_TIMEOUT_MS now bounds the RPC call alone, not the lock acquisition. Lock-contention during an in-flight prove queues patiently behind it (the prove yields the wasm-client mutex during the offscreen step via yieldWasmClientLock); only an actual unresponsive RPC trips the timeout. Three consecutive timeouts still open the circuit breaker, but they only fire when the network is really sick — not when the wallet is just busy proving.
3. Harness fixes — the test infrastructure changes that let CI catch this class of regression:
   • playwright/e2e/helpers/miden-cli.ts: add --locked to cargo install miden-client-cli. The published 0.14.8 crate ships a Cargo.lock pinning miden-assembly = 0.22.1, the same version the wallet's bundled SDK uses. Without --locked, cargo re-resolved to miden-assembly@0.22.3 which silently shifts BasicFungibleFaucet MAST procedure roots (see miden-vm#3144). The wallet then fails to recognize CLI-deployed faucets and renders their tokens as "Unknown", which any test selector filtering by token symbol can't match.
   • WalletPage.setDelegateProofEnabled(enabled) added to the WalletPage interface and all three impls (Chrome, iOS, Android). Just flips localStorage['delegate_proof_setting_key']; the wallet reads that synchronously per render so the next form interaction picks it up.
   • playwright/e2e/stress/stress.spec.ts: optional STRESS_LOCAL_PROVING=true env var that flips both stress wallets to local proving after creation. Lets you run the existing stress suite end-to-end against the local-prove pipeline.
   • playwright/e2e/tests/send-public-local-prove.spec.ts: focused regression spec — single send/claim with local proving forced. Smaller scope than the full stress, fast to run for targeted CI gating.
   • src/lib/miden/sdk/miden-client-interface.ts: keeps the per-step prove-timing markers I added to localize the deadlock (between executeTransaction / proveViaOffscreen / submitProvenTransaction / applyTransaction). +20 LoC of recordProveTiming calls; gated on MIDEN_E2E_TEST=true so production users see nothing.

How they fit

The SDK deadlock (#152) is what makes local proving impossible. Once unblocked, the next user-visible failure mode is the false "node unreachable" banner — that's #2 here. Together, local proving works AND looks healthy to the user.

The harness changes (#3) are independent of either fix's correctness — they just let CI exercise the path that was previously broken without anyone noticing.

Empirical verification

With this PR's harness changes + web-sdk#152's SDK fix linked in, ran the 5-tx stress against testnet with STRESS_LOCAL_PROVING=true:

Pre-fix: every consume hung at inner.getInputNote; zero sends completed; 5-minute Playwright timeout.

Post-fix:

• 5 of 5 sends completed end-to-end through the offscreen-doc prove pipeline
• [mt-offscreen-prove] tx_completed prove_ms=… fires for every op
• Send latencies: min 10.8 s, p50 15.7 s, p95 17.7 s (10–18 s/send = ~8 s pure prove + execute + submit + apply RPC overhead)
• Zero SyncManager false-positives during prove — banner stays clear

(Remaining stress harness failures trace to the conservation-check accounting drift and the getByText('TST') selector vs "Unknown" tile rendering — pre-existing, downstream of miden-vm#3144, unrelated to anything here. The --locked change in this PR closes the harness-side variant by pinning the CLI's transitive miden-assembly to the same version the wallet SDK bundles.)

Verification done locally

• yarn lint — clean
• yarn format — clean
• npx tsc --noEmit — clean
• yarn test — 1865/1865 pass
• yarn build — succeeds

Test plan

• Wait for linked-web-sdk-pr-ready status (web-sdk#152 merge + release)
• CI green on this branch
• Manual: enable Settings → General → Local proving, send a tx → completes in ~10 s (was: hang forever)
• Manual: same flow with delegated proving → unchanged (no regression on the delegate path)
• Stress harness rerun on testnet with STRESS_LOCAL_PROVING=true after merge