FLE Studio 0.3.2

• Speculation menu: Engine → Advanced disables the Speculation setting, with the reason underneath, when it would
  change nothing: with an upstream llama.cpp engine or a start profile. The plan answer
  (POST /api/engine/plan) carries speculation: {applies, reason} for this.
• Command card stays in view: in Engine → Advanced the Command card sticks to the top while the long settings
  column scrolls (own scroll when taller than the window; below 960 px it lies under the settings as before).
• Summary chips: the plan answer carries summary: [{label, value, tone}] (engine, effective speculation, MTP
  files, context, per slot, slots, fallback); Engine shows it as a chip row under the Command heading.
• Fallback is visible: the plan-time upstream fallback now says that the FLE plan could not be built and why (in
  the Command card, the Speculation hint and the status card). The reason of a load-failure fallback is the first line
  of the new run's log file, too.
• Models: every catalog card has an MTP badge (ready / files missing / no MTP); catalog entries carry
  mtp_status.
• Experiments: the comparison with the last run has a small history line of the decode speed of the last runs of
  the same setting; runs cut off at the token limit are grey ticks (series in the comparison).
• e2e: steps for the Speculation menu (upstream disabled with reason, FLE live preview, MTP download button, slot
  note); FLE_E2E_BROWSER runs the test with a Chromium binary of your own.
• e2e: the Forge step waits up to 45 s (was 15 s) for the published profile on slow macOS runners; the limit can be
  set with FLE_E2E_FORGE_WAIT.

Fixes

• The upstream fallback of the plan never ran from the page: the page sends the profile "auto", which was taken for
  a profile of its own. It now falls back as described for 0.3.1.
• Engine → Advanced: the opened "More: expert cache …" section no longer folds shut when the form is rebuilt.