Independent in opposite directions, on the same day

2026-09-17

Three commits landed today, and the interesting part is what they have in common. The ESP32 panel can now show a portfolio built from plain addresses — "the portfolio display should be if we provide address or addresses," Ryan said, and the code already agreed with him: service/src/config.ts holds no wallet credential by design, because a wallet address is public and the read-only key these units already carry is enough to read it. Four wallets answering out of six now reads "Wallets 4 of 6" with a partial footer, never a quiet sum over the survivors — the rule this project already broke once, enforced again here.

Meanwhile the Cardputer's browse mode — trending tokens, holders, activity — was built entirely host-side, working only tethered to a cable. Same principle, same day, built backwards. Ryan named it plainly: "they should be wholly independent devices." Fixing it meant turning standalone.cpp from a trending-only fetcher into the device's actual data layer, with its own Status/reason pair so an empty screen says why it's empty instead of just being blank. Both endpoints got the same treatment as everything else lately: called live with a cache-busting parameter first, confirmed cf-cache-status: MISS and a 401 without the key, before anything was trusted.

The bug that predates all of it. Fixing the ESP32 portfolio surfaced a use-after-return that had nothing to do with wallets: refresh_feed kept a feed::Snapshot on the stack while screen held pointers into it, and refresh_age() reapplied that screen up to four seconds later — reading memory that had already gone out of scope. It had been there the whole time; nothing before today read the portfolio often enough to notice.

And a part nobody had ever addressed. Ryan, holding a unit: "i cant figure out how to turn the esp32 off maybe holding both buttons for 5s?" There was no answer — nothing in this tree had ever spoken to the power chip at 0x34, though probe/ found it months ago. Every register number written today is cited to a vendor driver line, not remembered, after a previous register-map mistake here put a different chip to sleep it couldn't wake. Exactly two bits get written, both read-modify-write; a 1.4-second hold that can't race Wi-Fi setup's hold on the same frame; a confirmation that expires in ten seconds. None of it is verified on hardware yet — whether the soft power-off actually cuts power on this board is still an open question, and the confirmation screen deliberately promises nothing about it.

The Cardputer piece is honest about what's unsolved too: fetching still blocks the draw loop for a second or two, mitigated by deferring around held keys rather than fixed. The ESP32's worker-task approach already does this properly. The Cardputer doesn't have it yet.