Fix encoder readiness and bundle workflow authoring repairs

[sdevil7th]

Summary

Backend companion to the workflow-authoring fixes:

• Declare Encode Prompt's text-encoder connection as required, allowing current-registry preflight to catch missing dependencies in older saved workflows.
• Explain unqualified Auto plans with the requested model, precision and workload; installed weights do not establish Auto qualification. Keep the existing qualification checks unchanged.
• Document nested Block creation, model-switch interface protection and scalar input behavior.
• Bundle the paired MoDiff-client build, preserving backend-owned web/user/.

Client source commit: 73c659820441f262f7b348cf1d7c2b537ef9b0bc
Paired client PR: sdevil7th/MoDiff-client#6

Validation

• Full backend pytest: 3,724 passed, 546 skipped, 10,283 subtests passed.
• Ruff E9/F, uv pip check, POSIX launcher syntax, and runtime preflight with port 8088 free: passed.
• Client quality and browser results: final npm run check passed; shared controls 3/3; full mocked browser run 225/230 followed by all five corrected failures plus three related scenarios passing in an eight-test rerun. Production bundle cold startup had no JavaScript or asset-loading errors.
• Focused live outputs: saved Qwen Block switched to Z-Image, Save/reload, 512×512 Custom-memory execution and custom palette branch completed in 34 seconds; partial core grouping retained seven wires and completed Auto execution in 33 seconds after reconnect and reload.

Scope and compatibility

No executor, model-loading, package, file-access, HTTP or WebSocket contract is replaced. Required-input metadata is additive. Auto admission is not broadened. Existing saved Blocks remain readable; native creation in the bundled client writes V2.

The skipped backend cases retain their existing environment/optional-runtime conditions; no all-model or cross-platform output qualification is claimed. FLUX Auto remains refused without an accepted exact recipe, and SDXL prompt quality is outside these application fixes.

AI assistance: Codex implementation, diff review and verification. Human maintainer review remains pending. No credentials, private media or local runtime evidence are committed.