`[[deploy.data]]` is invisible: the gcu-c plate never mentions it, so agents write the esptool wall it replaced

[tig]

Summary

silico/partition_data.py exists, is wired into deploy_idf.py, has tests (tests/test_partition_data.py), and its own docstring states the intent plainly:

Data-partition assets for language=c deploy (no free-form esptool essay) … silico deploy plans these after the app image and flashes with esptool only after --yes (same confirm manners as the IDF image) — #79.

It appears nowhere in the gcu-c plate. grep -rn "deploy.data\|assets" silico/plates/gcu-c/ returns nothing — not in silico.toml, not in AGENTS.md, not in install/README.md.

What that cost on a real ship

tig/xuss-c streams a multi-MB PCM track from a SPIFFS partition. Its silico.toml has no [[deploy.data]], and install/README.md instructs the operator to hand-run exactly the wall #79 built the feature to delete:

py -3 tools/gen_spiffs_image.py --idf-path %IDF_PATH% --assets assets --out build/spiffs.bin --size 0xE70000
python -m esptool --chip esp32 -p COMx -b 460800 write_flash --flash_size 16MB 0x190000 build/spiffs.bin

That is a two-command wall in operator-facing docs, which Bedside 2 and 4 forbid, on a path silico deploy --verify cannot verify because it never placed the asset.

And it propagated into product behavior. Because the asset is flashed out-of-band, it can plausibly be absent — so the firmware grew a fallback that plays the boot riff in place of the missing song. The product spec (§4.4) requires the opposite: "refuse playback with a clear link status." A discoverability gap in the plate turned into a silent-degradation defect in the product.

Fix

Partly landed in #115 — commented [[deploy.data]] block in the plate's silico.toml and a note in install/README.md saying not to paste raw esptool write_flash into operator docs.

Still open:

• Audit the rest of the CLI surface the same way: which shipped features does no plate mention? partition_data was found by accident. board_profile, parts, product_path, and agents_stage are worth the same check.
• Consider a silico doctor hint when a language=c product has a non-app data partition in partitions.csv but no [[deploy.data]] — that combination is almost always a hand-rolled flash step.
• Consider whether silico deploy --verify should say out loud that it verified the image only when data assets are undeclared.

Note for whoever picks this up

The general lesson is worth more than the specific fix: on this spine a feature that the plate does not scaffold or name effectively does not exist, no matter how well it is implemented or tested. Shipping a CLI feature is not done until a plate references it.