feat(rtt): first-class RTT console support with eject path

[apage224]

Summary

RTT setup is currently a manual painpoint: users must locate the _SEGGER_RTT control block address, pick the right J-Link probe, and start/drain the buffer by hand. Meanwhile nsx view silently captures nothing for RTT-only apps (e.g. coremark), because nothing declares the app's console transport.

NSX already owns all the data needed to remove this friction (built ELF, SEGGER device/clock config, and --probe-serial plumbing), so RTT can become a first-class console transport that mirrors the existing flash/view model while preserving the ability to "eject" to raw SEGGER/pylink commands.

Motivation

Observed during the AP510 runtime sweep: coremark flashed successfully but nsx view showed no output because its score is emitted over SEGGER RTT (channel 0), not SWO/ITM. See examples/coremark/README.md and the working-but-manual helper examples/coremark/tools/rtt_capture.py, which requires the user to pass a hardcoded --rtt-addr and cannot select a probe serial.

Proposed work

1. nsx rtt command (mirrors nsx view)

• Same app/board selectors and --probe-serial plumbing already used by view_app (src/neuralspotx/api/_app.py) and the CLI (src/neuralspotx/cli/__init__.py).
• Auto-extract the _SEGGER_RTT address directly from the built ELF symbol table (no manual hex, no map parsing).
• Start RTT, wait for the up-buffer to come live (race already handled in examples/coremark/tools/rtt_capture.py), then stream channel 0.
• Support --out/--file line-streamed capture, matching view --file.

2. Declare console transport in nsx.yml

console:
  transport: rtt   # or swo (default)
  rtt_channel: 0

• nsx view on an RTT app prints a friendly redirect to nsx rtt instead of silently capturing nothing.
• nsx rtt on a pure-SWO app warns symmetrically.
• Removes README/tribal-knowledge as the only source of transport truth.

3. Preserve the eject path

• --print-cmd / --dry-run: emit the fully resolved invocation (device, speed, probe serial, block address) without running.
• --print-address: print just the resolved _SEGGER_RTT address (the one annoying-to-find value).
• Generate a committed per-app/board RTT artifact under build/<board>/jlink/<app>/, mirroring how flash_cmds.jlink / reset_cmds.jlink are generated in src/neuralspotx/cmake/nsx_helpers.cmake.

Rollout order

1. nsx rtt with auto-address extraction + probe-serial reuse (self-contained, biggest win).
2. console.transport in nsx.yml + cross-warnings in view/rtt.
3. Eject affordances (--print-cmd, --print-address, generated RTT artifact).

Acceptance criteria

• nsx rtt <app> --board <board> --probe-serial <sn> captures coremark output on AP510 with no manual address.
• nsx view on an RTT-declared app redirects to nsx rtt instead of empty capture.
• Resolved RTT command/address are inspectable for the manual SEGGER/pylink workflow.

References

• examples/coremark/README.md (RTT vs SWO note)
• examples/coremark/tools/rtt_capture.py (current manual helper)
• src/neuralspotx/cmake/nsx_helpers.cmake (flash/view/reset target generation)
• src/neuralspotx/api/_app.py, src/neuralspotx/cli/__init__.py (probe-serial plumbing)

Follow-up PR to implement.