[Strycher]

E4 band 4, task #265, Epic #212. Opened autonomously under the owner's standing grant of 2026-09-07 — and deliberately not merged under it: #265's acceptance requires full hardware verification, every screen and every peripheral, before merge. Nothing in E4 has been flashed. This PR waits for that.

What moved — three proven commits on one branch

commit	unit(s)	functions
1	lvgl_port — the five LVGL callbacks, initFCTheme, the display/indev objects, the touch statistics	6
2	oled_screens — updateOLED and the five drawOLED*	6
3	screen_compass, screen_env, screen_telemetry, screen_geocache, screen_settings, navigation — every widget of every screen, the navigate functions, fcHeaderCreate, the list picker, updateSDIndicators, updateDisplay	58

Commit 3 is six units in one step because the screens and the navigation reference each other through headers; any one alone would have reached back into src.ino. initLVGL stays as the composition root. A fourth commit is the residue clean-up and the record.

#265's acceptance, measured

• src/src.ino: 750 lines — setup, loop, initLVGL, initSD, scanI2C, handleButtons and its two helpers; 6 file-scope variables.
• python scripts/reference-graph.py: 7 of 8 functions touch no shared state of the file, the eighth only its own three; cut set 0, largest cluster 1.
• python scripts/symbol-inventory.py: 8 functions, 6 variables — the residue only.
• Hardware verification: yours. Every screen and every peripheral, per E4-6 (band 4): move the consumers and screens last #265, then merge.

Calls made under the grant — please review

1. Band 4 as one branch and one open PR rather than five merged children, because of the merge gate above.
2. initLVGL stays in src.ino (it builds every screen; composition root); lvglTickCb lost static because it registers it.
3. web.h gained using fs::FS; — TFT_eSPI defines FS_NO_GLOBALS, and src.ino's include order had hidden that WebServer.h depends on the alias.
4. Six statics the graph could not see (write-only headers, const-folded counts, an inlined helper, the settings menu buttons) were moved by hand and are marked in their units.
5. The LVGL log callback's LV_USE_LOG guard was restored by hand; the harness carries functions, not the preprocessor lines around them.

Proof

scripts/verify_extraction.py PASS on each commit; full figures in docs/god-object-inventory.md §15:

unit 1  592 -> 592 symbols, 27 relocated, 13 linkage changes declared, 5 deltas net +2      Flash -20
unit 2  592 -> 592, 6 relocated, 1 delta, 6 moved checked                                   Flash +52
unit 3  592 -> 593 (one compiler clone, named), 213 relocated, 12 linkage changes declared,
        6 deltas net +104, 61 moved functions checked by reference set                      Flash +88
Flash 1,856,690 -> 1,856,810; RAM 103,248 unchanged. Both toolchains build. 58 tests pass.

E4 in total

src/src.ino 8,742 → 750 lines into 25 translation units and 28 headers; Flash +848 bytes (0.05 %) with every byte traced; RAM +16. Every move proven behaviour-neutral by symbol set and, from band 3c, by every moved function's reference set. The two harness faults the dry runs caught this band are in §15.

[Strycher]

Hardware verification evidence for band 4 arrived through PR #288, whose branch is this PR's head plus the UM FeatherS3D env and two fixes (#284 TFT_eSPI SPI port, #285 watchdog). On um_feather at 2026-09-08 02:16 EDT every unit that band 4 extracted was exercised from the boot log: lvgl_port, the five screens, navigation, oled_screens; TFT drawing, touch found and armed, OLED up, and the loop at 44 Hz with the owner confirming the TFT sleep timeout. Nothing in band 4 misbehaved; the two defects found were pre-existing PlatformIO-build defects that no image had met before, fixed in #288. Merge order: this PR first, then #288 retargets to main.