Return Fire: retained-image review still gates native calibration
Run: scratch/runs/20261109-return-fire-reliable-review-jeep-entry-2105/.
Inspected checkout: b214d6da741ed9375d81683c338a9ddd00aecc0d.
Reliable model presentation was not established, including with an enlarged game-window crop, an independently encoded PNG, JPEG and an explicit PNG image content block. No native fixture or original was started. Bunker entry, jeep deployment, flag pickup and return remain unverified. No emulator defect or production repair is claimed.
Read corrected bunker-entry2045, blank-transition2055 and armored-flag1940. The earlier supposed guest disappearance remains disproved by the retained capture bytes. This run extends that investigation to smaller, independently verified presentations before spending a new native allocation.
Retained bytes and alternate presentations
The copied historical command5 PNG remains 25,334 bytes with SHA-256
7b27f488f0624cbde33f68f3938a24606eef4208c01d755d88e3730143388ae3.
Node pngjs and a retained Python standard-library CRC/zlib/unfilter decoder
independently yield RGBA SHA-256
636aefff98a657423518f12fb8686b4f87bdbb15ba9994ee59adb49fd3517912.
It is 945×889, with 91 colors and 91,564 non-teal pixels in the game rectangle.
audit.cjs checks every pixel of four enlarged presentations against its source
pixel: two independently encoded 2× game-window crops, a 3× movement panel and a
3× dashboard crop. The two full-window crops have different encoded hashes but
identical decoded RGBA SHA-256
5c19bf8c6918698854166b721ff0386ba21dcbb1b536ab973ce20dd586e55ab7.
The tool-returned movement-panel data URL was also compared directly with the
file's base64; it matches exactly. This verifies byte preservation, not model
perception. No source or media archive was made.
Nine separate model-review calls are recorded in
presentation-observations.json. Several clearly showed the window, jeep,
ammunition and dashboard. Repeated calls were sometimes incomplete or ambiguous
to the model, including the independent crop, JPEG and explicit image/png
movement panel. Ambiguous calls did not authorize dependent input. This run
does not establish a platform mechanism, caching cause, or MIME-only repair.
Numerical decoding cannot substitute for reliable review of a live terrain
frame, so no native execution followed.
Controls and the next calibration
Read the actual local RFIRE.HLP using both ordinary strings and one-character
minimum strings, retaining short excerpts and its SHA-256
ec9679b23db09803498b5292b22b35c84a1d31d2a62949983532dc26017d43c7.
It specifies H/Enter deployment, J map, K information; per-vehicle Left/Right
movement-mode selection while viewing status; W forward, S reverse, A/D turns;
and entry by driving over the bunker until its light comes on, then firing.
The jeep's heading beacon is separately documented and cannot establish entry.
The reviewed icon has a bright straight-arrow left panel and dark bent-arrow right panel. Interpreting these as vehicle-relative and screen-relative is an icon inference; field displacement remains unmeasured. The next plan therefore preserves the selected default mode and uses the documented W/S movement and A/D turns for short, reviewed near-home calibration. It does not repeat Right100ms or assume a successful toggle. Close information and review actual selection before H deployment. Verify the entry light and underground return transition before an armored exposure route or jeep retrieval attempt.
next-calibration-plan.json retains that sequence and the prior source-derived
A1 route. The historically exposed flag is not assumed to persist into a fresh
original process. All native allocation, immutable watchdog, review and terminal
failure requirements remain unchanged.
Validation and lifecycle
Exact image/payload audit passes. Tracked composition: 33 cases; collector: 129;
native keys: five; lifecycle: six; locked build: pass. The first lifecycle test
run read an empty launcher-result.json between its creation and write; the
fixture writes it non-atomically. Its completed file and expired-unbound fake
watchdog are retained. A fresh offline invocation passed all six scenarios;
no test, provider, deadline or production code was changed to obtain that pass.
An unavailable optional Python Pillow import is retained in invocation history;
the actual independent decoder uses only the standard library.
Source identity is the exact Git SHA plus the small relevant patch. Image variants, decoder, audit and logs are independently reread for SHA/size, with PNG CRC and decompression checks. No shared pinned providers or Q3/Nox startup files/tests were edited.
No box was created, leased, resumed, captured, stopped or otherwise controlled.
New box ID, allocation UUID, state, expiry, stop receipt and actual watchdog
join are not applicable. Lifecycle tests use fake API adapters only. The
historical Return Fire box remains historical evidence; this task did not
alter it. Free disk on /home remained above 1.5 GB.
The concrete blocker is repeatable model review after correct retained-byte delivery and decoding. Resolve that boundary before native near-home entry calibration; there is no justified emulator patch from these observations.