Add sensory action bar and transparent neural inspector

[bitdeep]

Trial controls were buried in a text-heavy neural panel, obscuring the animal.
This change puts Feed, Water, Bitter and Antenna in a persistent action bar and
turns Neural activity into a compact transparent inspector. Hover, focus and
click/tap tooltips explain components and measured values; graph position selects
a recorded sample. Baseline, block and Stop apply to the selected protocol.

Water and bitter use the published sensory IDs extracted from the existing
checksum-verified notebook and the same complete FlyWire graph. Sugar and water
can drive the experimental MN9-to-rostrum adapter. Bitter alone produces neural
activity without MN9 spikes or motion. Wings and Walk remain explicitly
unconnected; Feed represents feeding initiation, not eating or digestion.

Validation:

• Seven physical/transport checks passed in bounded Docker tooling.
• Full-graph causal comparisons passed: sugar MN9 30/17 and 40.03° movement;
  water 14/7 and 29.96°; bitter 0/0 with 677 downstream spikes and zero movement.
  Blocked links preserve neural counts with zero actuator force.
• Shared clocks, excluded actuators, invalid preset rejection, motor cancellation
  and antennal stop/restart passed; the local report binds current code hashes.
• Browser checks covered all four presets, commands with the inspector closed,
  measured trace tooltips, keyboard focus, Escape and 390×844 layout. No horizontal
  overflow, JavaScript exceptions or failed requests were observed. A software
  graphics context reset on resize recovered.
• JavaScript/Python syntax passed. No added dependencies, continuous render loop,
  backend GPU allocation or resource-limit increase. A resting-service sample
  after closing the test browser used 146.5 MiB and 7.82% of one CPU core.

README, browser controls, source/protocol documentation and measured evidence are
updated together. These checks establish a software-to-actuator path, not
biological equivalence or complete embodied locomotion.