fix(desktop): full-jitter backoff on gateway WS reconnect (salvage #76282)

[kshitijk4poor]

Context

Every desktop client that loses its gateway WebSocket redials on the same deterministic ladder (1s·2^n capped at 15s in two places, 30s in a third). A gateway restart makes every open desktop window redial in lockstep — the thundering-herd shape full jitter exists to break. WHO: anyone running multiple desktop windows/machines against one gateway, most visibly during gateway restarts/updates.

This centralizes the scattered backoff policy into one resolver (lib/reconnect-backoff.ts, AWS full-jitter: random[0, min(cap, base·2^n))) and applies it to all three reconnect sites — matching the desktop guide's "one resolver owns each policy" rule.

Provenance and the escalation fix

Salvage of #76282 (author trevornk; attribution mapping merged as #77641). Stage-1 review found a real blocker in the original: RECONNECT_ESCALATE_AFTER = 6 counted ATTEMPTS, calibrated to the old deterministic ladder (~45s of failures before surfacing "Lost connection"). Full jitter reaches 6 attempts in ~9s expected — so a brief network blip would flash a boot error. An existing test caught it deterministically. The follow-up commit switches escalation to elapsed time since first failure (constant calibrated to the old ladder's wall-clock behavior), making that existing test pass UNCHANGED — the test is the contract.

Measured impact

Behavioral: jitter bounds verified [0, min(cap, base·2^n)), no overflow; escalation timing preserved at the old ladder's wall-clock calibration. Herd-dispersion is probabilistic by construction (full jitter is the AWS-recommended variant for exactly this).

Verification

• Full use-gateway-boot.test.tsx suite + store/gateway tests green on the final stack, including the post-boot-blip contract test that failed on the unfixed pick.
• Mutation checks: de-jittering the backoff fails the jitter test; reverting the escalation fix makes the post-boot-blip test fail again; restore → green.

Closes #76282.