feat(wifi/lan): WiFi/LAN transport + tinfl inflate; pin pioarduino 55.03.39

[davelee98]

Ports the WiFi/LAN transport and the ROM-tinfl inflate path onto current main, plus the build and cleanup work that came out of getting it green.

Five commits, each self-contained — happy to split the build/refactor ones into separate PRs if you'd rather review the feature alone.

2f38224 — WiFi/LAN transport + tinfl inflate

Squashed replay of a 31-commit branch. A commit-by-commit rebase wasn't viable: #120 (Seeed GFX → FastEPD) and #121 (writeSerial → od_log_*) touch the same files, so every intermediate commit conflicted on the same renames and none would have compiled.

• LAN transport — mDNS identity TXT, port/mode selection, TLS-PSK listener, command-origin routing with CCM bypass on TLS, per-transport session ownership, Nagle off, drain-to-dry RX with single compaction.
• WiFi — strongest-AP selection with RTC-cached BSSID, roaming at −75 dBm, all work moved out of the event callbacks (fixes a panic — handlers ran String concatenation on a 4 KB stack), lazy RSSI-low handler registration.
• TLS — both mbedTLS record buffers pre-reserved at boot while internal DRAM is still contiguous. ssl_setup was failing in the field at 51 KB free / 31.7 KB largest block, because it needs two ~16.7 KB contiguous internal-DRAM blocks and cycles 33 KB per connection.
• Inflate — compressed transfers routed through ROM tinfl, LZ dict sized to the actual DEFLATE window.

Conflict resolutions kept your od_log_* form while preserving this branch's content: transport tag and pre-enqueue [Q:n] depth on the TX line, backlog-only queue logging, compression mode/ratio on DW start/complete, and SSID/password redaction in the config dump. setConfigLoggingQuiet()/cfgInfo() were dropped — they existed to skip the config dump in the deep-sleep wake window, which the compile-time OD_LOG_LEVEL default (INFO) now does by eliding every od_log_debug.

Note: the vendored include/opendisplay_protocol.h resolved to main's copy, which is newer than the canonical one in opendisplay-protocol (upstream added the 0x43 version-response docs). That repo needs a --pull, or the next --push reverts them.

2645c43 — pin pioarduino 55.03.39, make the C6 mempool script non-fatal

Moves the pin from 55.03.32 forward to 55.03.39 (Arduino 3.3.9 / IDF 5.5.4). This restores the platform's own merged-image post-build action — firmware.factory.bin next to firmware.bin, flashable at offset 0. That action doesn't exist in 55.03.32, so local merged images disappeared when the floating stable URL was pinned back in 0d95b37.

Deliberately not 55.03.311, which is what stable currently serves. IDF 5.5.5 drops the r_ prefix on the C6 BLE controller's mempool exports and moves them from libble_app.a to libbt.a (r_os_mempool_init → os_mempool_init, likewise os_memblock_get/_put). NimBLE-Arduino 2.5.0 — latest release, 2026-04-02 — still calls the r_ names, so esp32-c6-N4 fails to link there. No library-side fix exists; it has to land upstream in NimBLE-Arduino. Only C6 is affected: it alone ships a precompiled controller blob carrying its own NimBLE porting layer (CONFIG_BT_LE_CONTROLLER_NPL_OS_PORTING_SUPPORT=y). Full analysis in docs/FINDINGS_C6_NIMBLE_IDF555_MEMPOOL_ABI_2026-07-25.md, and the constraint is recorded in platformio.ini and CLAUDE.md so the next bump doesn't walk into it.

esp32c6_nimble_mempool_link.py now warns and skips on every failure path instead of sys.exit(1). On 55.03.311 the hard exit aborted with no entry os_mempool.c.o in archive, hiding the real cause behind a missing-member error; skipping lets the link proceed and report the genuine undefined reference to r_os_mempool_init.

Measured across three platform versions, all eleven envs:

Platform	C6 without script	C6 with script	Other 10
55.03.32	links	links (no-op, +28 B)	OK
55.03.39	links	links (no-op, +148 B)	OK
55.03.311	link fails	link fails	OK

47432e0 — esp32-s3-N16R8-extuart-debug env

OD_LOG_LEVEL defaults to INFO, which compiles away every od_log_debug() — and that's where the config dump, BLE RX/TX hex with [BLE][Q:n] depth, and DIRECT_WRITE throughput live. Adds a variant env via extends (same pattern as esp32-s3-E1004) rather than a flag on the shipping one. Costs 12,392 bytes flash, no extra RAM. CI and release.yml enumerate envs explicitly, so neither picks it up.

799f324 — delete the WiFi power-save scaffolding

The transfer-scoped WIFI_PS_NONE experiment was already gutted after hardware measurement showed it collapses throughput under BLE coex. What remained was a subsystem that couldn't do anything: lanPsSuspended was never assigned true anywhere, so lanPowerSaveSuspend() was an empty body, lanPowerSaveRestore() early-returned before touching the radio, and lanPowerSaveSuspended() was a constant false — yet eight call sites and an 18-line loop() safety net with its own timer and warning log still guarded it, and the header advertised the original semantics with no hint the functions were inert.

Also drops the two esp_wifi_set_ps(WIFI_PS_MIN_MODEM) calls, so the firmware never sets power-save mode at all. MIN_MODEM is the driver default and exactly what those calls set, so runtime behaviour should be unchanged — flagging it explicitly since that's an assumption about the IDF default rather than something measured. A note in wifi_service.h records why nothing calls esp_wifi_set_ps(), so "force PS_NONE during transfers" doesn't read like an untried optimisation.

Separately: the two DW complete: … KB/s lines move from od_log_debug to od_log_info, so upload throughput is visible in a normal build.

d785ac0 — transferActive()

Three flags mark an in-flight transfer — directWriteActive (global in main.h), pipeState.active and partialCtx.active (file-static in display_service.cpp). Callers meaning "a transfer is in flight" spelled the disjunction out themselves, and drifted: the log predicates had all three (longhand, twice), the roam gate had direct+pipe, the touch poll had direct only.

Two behaviour changes fall out, both widening an existing skip:

• Roam gate — a BLE-origin PARTIAL write with no LAN client attached could be interrupted mid-stream by a full-channel scan. Neither the LAN-client check nor the direct+pipe check fired, though the surrounding comment already claimed BLE transfers were protected.
• Touch poll — GT911 I²C reads are now skipped during PIPE and PARTIAL too. The original rationale (bus and loop-task contention) applies identically to all three.

Clear-timing was checked before unifying: all three flags are cleared before their panel teardown, so the predicate drops at the same relative point and gates open no earlier than before. Left deliberately specific: the direct-write watchdog and teardown in main.cpp, and the (directWriteActive || partialCtx.active) 0x0072 session-ownership check that excludes PIPE.

Also removes two dead symbols: handleButtonPress(uint8_t) (declared 2026-02-12, never defined or called — processButtonEvents() does the job) and partialWriteActive() (only caller was the deleted power-save net).

Verification

Hardware: the WiFi/LAN transport and TLS-PSK encryption have been flashed and exercised end-to-end on an E1003 — the full chain from Home Assistant through py-opendisplay to the firmware, not just a local loopback. That covers the substance of 2f38224: association, mDNS discovery, the TLS-PSK listener, command-origin routing with the CCM bypass, and image delivery over LAN.

Build: all eleven environments compile on 55.03.39 — nrf52840custom, all five S3 variants, esp32-s3-E1004, both C3, esp32-c6-N4, esp32-N4.

Not separately exercised on hardware, and worth a look during review:

• C6 BLE on the newer IDF. The pin bump changes the controller blob under NimBLE-Arduino; C6 links and builds clean, but a C6 board should advertise, connect and complete a transfer before this is trusted.
• The two gate changes in d785ac0 — roaming now deferred during PARTIAL writes, touch polling suppressed during PIPE/PARTIAL. Both only widen an existing skip, so the worst case should be touch being unresponsive slightly longer during an upload; a GT911 device running a long partial-write stream is the case to exercise.
• The esp_wifi_set_ps() removal rests on WIFI_PS_MIN_MODEM being the driver default, which is reasoning from the IDF default rather than a measurement.

One environment note: 55.03.311 requires PlatformIO Core ≥ 6.1.19 (it silently uninstalls itself on older Core). The pinned .39 has no such requirement.