Convoy route investigation: 04_009 / 7 (level 80)

Objects 6–17 fan out onto translated paths rather than following one convoy route. Retail playtesting confirms both the unusual movement and vehicles driving through solid buildings, sometimes drawing over them. This is an original-game defect, not evidence that the route decoder needs road snapping. Complete trajectories and predicted off-map endpoints remain unverified. The viewer retains the decoded routes and identifies the remaining uncertainty.

What was checked

The reproducible audit checks:

• Original BE07 setup constructs all 17 routed vehicles with the exported positions, headings and route pointers. Their initial repeat state is zero and time field is F000h, matching the preview's initialisation.
• The original image and saved DOSBox memory capture agree byte-for-byte across 8F3B–924F, covering the interpreter, action table and movement/effect handlers. This excludes a mismatch against that captured version, not later mission-specific changes.
• Sampling at 256 and 512 ticks gives the same first vehicle's operation sequence and endpoint. This is not an arrow-density or coarse-sampling problem.
• The existing Wrecker and first-flatbed checks still pass; they do not validate this convoy's live behaviour.

At resource offset 42B5, the source stream begins:

C1 B1 A1 71 61 51 41 31 21 11 01 F1 03 ...

Vehicles 6–17 enter at successive bytes in this prefix. Their starting rows are 62, 63, 64, 67, 68, 69, 70, 71, 72, 73, 74 and 75, all in column 22 heading south. In isolated execution, vehicle 6 consumes C1 for 13 cells, then B1 for another 12, then A1 for another 11, and so on. The first three fetches occur at ticks 0, 53,248 and 102,400. Its final predicted position is (62.5,159), outside the city. Later vehicles skip increasing portions of that cumulative prefix, hence the fan of translated routes.

The data has a suggestive pattern: each vehicle's initial forward count alone brings it to row 75, and F1 would then move all of them through a common 16-cell section. This is evidence of a possible intended convergence, not evidence that we should skip the intermediate bytes. The inspected native interpreter advances its saved pointer one byte per fetched instruction and does not perform that skip. Introducing one in the exporter would currently be an unsupported correction.

Vehicles 18–22 use other shared entries and converge to the same predicted endpoint, approximately (66.496,95.996). That does not independently prove their live paths, but localises the conspicuous discrepancy to the earlier shared prefix.

Live DOSBox checks (5 October 2026)

Four controlled starts reached this mission in a temporary game copy. The capture notes describe the debugger interventions and failures; the decoded dumps retain the object fields and memory hashes.

• At the first vehicle callback, the loaded prefix is still C1 B1 A1 ... and 8F3B–924F still matches the unpacked executable. The first truck is at (22.5,62.5), heading 8000h, with next-route pointer F136h, action zero and phase start F000h.
• The first fetch advances its route pointer to F137h (resource 42B6), consistent with reading C1. Captures at 8F78 precede storing the fetched action, so their action byte is still zero, not a completed run.
• Attempts then became unstable, logging illegal instructions or unhandled interrupt 1 repeatedly. Removing memory watchpoints did not prevent this. No reliable second fetch or long-running vehicle position was captured.

This establishes that the suspicious prefix survives loading and is entered as predicted. It does not establish that B1 is subsequently executed, or that any vehicle actually leaves the map. The startup shortcuts are also an unresolved source of interference; these failures are not evidence of a bug in normal gameplay.

Retail playtest follow-up (8 October 2026)

We observed the level-80 convoy in the original game and confirmed that its initial movement follows the unusual plotted routes. Further playtesting confirmed that the vehicles drive through solid buildings and can remain visible through them because of drawing order. The bad route therefore affects both physical placement and visible overlap in retail.

Together with the shared-prefix pattern, this strongly suggests an authoring mistake: the staggered vehicles were intended to join a single convoy path, but instead accumulate the preceding vehicles' forward runs. That intended path is an inference; the observed building traversal and overdraw are confirmed.

The faithful reconstruction must retain the original route interpreter and data, including these defects. Any corrected convoy route belongs to a separately labelled optional fix, not the baseline or the analysis route preview.

Behaviour after leaving the map area has not yet been confirmed. These observations do not validate every plotted segment or predicted endpoint; the viewer warning now distinguishes those limits from the confirmed retail bug.

Previous native checks established behaviour of an isolated interpreter with stated substitutions. They were insufficient evidence to call every plotted route faithful to the complete game.

PYTHONPATH=/tmp/darker-python python3 tools/audit_convoy_routes.py