sots-re/findings/subsystems/nav-classifier-live.md

The route classifier under a hook — game/nav from modelled to compared

Lane DN, 2026-09-09. Guest VM146 (sots-re-win10-146), build box CT111, own build dir /srv/re-lab/build/sots-engine-dn, own worktree sots-engine branch wip/dn.

Hold announced: VM146 only, for the duration of this lane. VM140 is the exclusive reference guest and was not touched; VM141 (lane CZ) and VM145 (lane CR) are held by other lanes and were not touched. VM146 is released restored and hash-verified at the end of this document.

game/nav is the campaign's largest modelled module by span of behaviour and it had never executed under an instrument. tools/displacement.py records it as

offline vs 58 waypoints / 46 flight plans across 11 saves, 0 failures — never instrumented; not one leg has executed under a hook

and path-solver.md closes with the same sentence from the other side: "the classifier's rules are read instruction by instruction … and not one leg of it has ever been observed executing under an instrument." This lane executes it.

---

0. The entry point is a real function, so the hook is on the classifier itself

The brief allowed for the classifier being inlined into its caller, in which case the honest hook would have been on the enclosing function comparing its output. It is not inlined. path-solver.md §0.1/§1.1/§2 read it from the instruction stream:

StarFleet_ClassifyLeg	0x00703730, thiscall, ret 0x14, six arguments, real body 0x00703730..0x00703bc9 then 6 × int3 to the next function start
callers	PathSolver 0x007066c0 at two sites (the pre-flight probe and the per-leg loop), and MoveFleet's multi-waypoint continuation at 0x7da5c6 / 0x7da5da
output surface	the return value (the waypoint kind), *flagsOut (the flag word), and the route record's three written words at routeOut+4..+0xf. Nothing else.

So the comparison surface is the classifier's own three outputs plus its return value, and there is no unit boundary invented here that the binary does not have. PathSolver is hooked too, but trace only — see §1.3.

---

1. Predictions (committed before the compare run, rule 2)

This section is committed in its own commit, before the instrumented build ran once.

1.1 The workload, and why it reaches two route classes

Save ar-oracle-A-pre.sav (15b99255…), turn 38, players: Human ×2 (net 16 = the human seat, 496), Tarkas ×2 (32, 512), NPC ×4 (528–576). DriveTypeOfSpecies maps Human → 3, Tarkas → 1, NPC → 0, so three drive classes are present before the AI does anything.

Procedure, fixed and replayed identically in every run:

1. load (SavedGames reset to exactly one file, row 1), lobby, Launch;
2. Move, order Delta Fleet → Isis. The game answers with its own prompt — "Delta Fleet does not have the range to reach Isis using nodes. Do you want to plot a sublight course?" — and OK commits it. This is the sublight-crawl class.
3. Move, order Delta Fleet → Venkman. Hover reads "Range: 6.0 / ETA: 2 Turns", there is no prompt, the order is committed straight away. This is the node-route class.
4. End Turn.

The hover readout named the class before either click, exactly as the campaign's guest notes say: Venkman ETA: 2 Turns (node), Isis ETA: 50 Turns in red (sublight); Nu Ophiuchi 4 Turns, Kua'Taak 47, Kam'Tara 53, Kao'Lolto 50, Ko'Kotan 47. The state was built deliberately from that readout rather than hoped for.

1.2 Falsifiable predictions

P1 — call count from the two player orders: 8. Each accepted order costs two PathSolver calls (the UI's dry run FUN_005e6d50, then the server's OrderFleetMove), and each PathSolver call on a one-destination list costs two ClassifyLeg calls (the pre-flight probe against MinRange, then the loop leg). Two orders × 2 × 2 = 8, with 4 PathSolver records. count is 1 on all four and the leading-destination drop does not fire (neither destination is the fleet or the fleet's own system). Falsified by: a different multiple of 2 — most likely because hovering also runs the dry run, which would add calls in pairs and is worth knowing either way.

P2 — the Isis leg raises 0x002, not 0x020. The game's own prompt says the fleet cannot reach Isis "using nodes", which is errBits seeded at 0x00703a0d and ORed at 0x00703b10: a node line exists and is beyond the fuel budget. So on that call line_found >= 0, the returned kind is 0, and the flag word is 0x002. Falsified by: line_found == -1 on that call, which would make the flag 0x020 and the prompt string generic rather than diagnostic.

P3 — the Venkman leg returns kind 3 with a live route. line_found >= 0, return value 3, nrp == line_found, nrf == Epsilon Eridani's network id, nrt == Venkman's, flags 0.

P4 — no Human or Zuul call returns 1, 2 or 6; no Tarkas call returns 3; every Tarkas call leaves the route record at {-1,0,0} and performs no range check. This is DriveTypeOfSpecies and step E of the decision order, live. The record carries the original's own StarFleet_GetDriveType next to the species, so the table is checkable rather than assumed.

P5 — can_bore == 0 on every call, so the bore arm is reached 0 times. No Zuul player, and the 0x20000 capability is the Zuul node cruiser. This is the declared gap that makes itself inert, and the trace proves it instead of the brief asserting it (rule 28 practice 1).

P6 — gate_cap == 0 on every call, so no call returns kind 4 or 5 and flag 0x004 never appears. Every player in every corpus save has NGts == 0. Kinds 4 and 5 stay at zero observations and this lane says so rather than quietly counting a clean sheet as coverage.

P7 — 0 divergences, and the reason is the narrowings already in the module. The MoveFleet lane's eight live divergences were all float32 narrowings; game/nav already carries the five of LegLength and the unstored r*r of LegInRange. I predict the compare comes back clean, and I predict that a boundary fuel case is the only thing in this subsystem that could break it — which this workload is not built to hit. Falsification shapes, each with a distinct symptom:

• kind differs on a Tarkas/NPC call → the drive table is wrong (visible directly against drive_orig);
• flags differ on 0x001 / 0x800 → the pending-action-mask split is wrong;
• flags differ between 0x002 and 0x020 → LegInRange disagrees off a boundary, i.e. the range or the tank-capacity input is wrong, not the arithmetic;
• nrf / nrt differ → the origin/destination identification in the node arm is wrong;
• the fleet guard fires → the classifier writes the moving fleet and the module's purity claim is false.

P8 — neutrality holds. The compare run's (Autosave EndTurn) and (Autosave) are byte-identical to the control's. If they are not, every number in this lane is void (rule 19).

P9 — at least 8 distinct input states. A count of calls is not coverage (rule 23); this is the number the verdict has to stand on.

1.3 What is deliberately not compared

• PathSolver is trace only, and its ours throws if anyone sets it to compare. The plan needs the fleet's effective range, which comes from FUN_00705d60's tanker-redistribution fold — a function path-solver.md §9 explicitly did not read. A hook that returned "no plan" there would print a green verdict over a missing input, which is rule 1's failure mode.
• nrp on an existing line is an input, not a result. The single-hop answer is taken from the original's own NodeGraph_FindNodeLine before the original runs (afterwards it could return a line the original itself had just bored). Its agreement is therefore not evidence; the kind, the flags and nrf/nrt are.
• Three input gaps, counted per call rather than argued away: the intercept solve for a destination fleet already on a node leg (intercept_gap), the node bore (bore_arm), and the unread friendliness test FUN_00817890 on the point↔system arms (friendly_recon).