Take the fishing cast curve from the game, and lead the fish instead of aiming at it

[averagenative]

Two commits on the fishing helper, 2.5 → 2.7 (suite 1.48 → 1.57). The complaint
that started it: "it's not actually on the line — if we go over the line we
tend to miss, but if we go under we can often catch." Both halves turned out to
be true, and neither was going to be fixed by fitting the curve harder.

The catch window had been in the code for two versions and never once drew

v2.4 added the tolerance to the species table and drew a bar as wide as it. The
field was never copied onto the detected fish, so the draw read undefined,
(f.catchN || 0) * laneW came out 0, and both the lane bar and the gauge band
collapsed to nothing every frame. It looked exactly like a helper that only
draws a line, because that is what it was.

The cast curve is no longer fitted

The minigame was pulled out of the client and reimplemented offline
(scripts.ActorEvents_229._event_Minigames1), and the cast is a dozen lines:
the hold advances an angle one degree every 10 ms update from 90, the gauge
shows 1 - |sin(angle)|, the release launches at (0.4 + 2.06p, -1.4 - 1.45p)
under gravity 0.05, and the update that would carry the bobber past the water
does not move it at all.

Transcribed, it reproduces the offline module's landings to 4e-13 of a lane
unit over all 91 casts, and sits 0.58% of the lane from the parabola fitted to
19 measured casts — inside that parabola's own 1.1% residual. The parabola's
error was never noise, it was shape: −1.9% of the lane through the middle,
against a fish window only ±5% wide. No refit removes that.

Calibration now learns two numbers (where the lane starts, how wide it is)
instead of three. Not a calVer bump — a sample is still a gauge reading
paired with a landing, so they are kept and refitted.

Two things fall out that a fit could not give. The gauge is a ladder of 91
rungs, unevenly spaced (0.8 lane units per rung at 6% fill, 6.6 at full), so a
catch window is countable: 30 to 210 ms depending on species and distance. And a
release 30 ms late lands 3 lane units further out at the bottom of the gauge
and 18 at the top, against windows of 15–23 — which is the whole of "aim at
the mark and I sail past it". tuning > Release lead carries that; the status
line measures what is left of it and refuses to guess when a cast cannot be
attributed to a single mark.

The fish do not hold still

From cast 6 the lane bobs and from 17 it widens. A cast is in the air 600–1160
ms, so a fish can cover up to 35 lane units while the bobber flies — more
than a whale's entire catch window.

mark placed at	casts called	really catch
where the fish is now	450	53%
where it will be on landing	437	99%

53% is a coin flip, and it is what every version before this drew. Each rung is
airborne for a different length of time, so there is no single position of the
fish to aim at: every rung is now tested against its own landing moment, and
what is drawn is the set of casts that actually connect, as bands of contiguous
rungs rather than one span.

The band goes where the fish will be, never across the stretch it drifts over —
a bar spanning the drift is mostly wrong at any moment and is the opposite of
the question. A separate dotted hairline, no fill, joins the band to the sprite,
because at a median 15 and worst 29 lane units of offset the band sits nowhere
near the fish it belongs to.

The fit needs time, not samples (0.5 s of history predicts 900 ms out with a
worst case of 16 lane units; 1.2 s with 4.5), and the residual cannot tell a
good fit from a bad one — fits that went on to miss by more than 6 units had an
rms of 0.28, against 0.30 for the ones that did not. So the gate is the span and
nothing else. Below it the band draws dashed and says it is not tracking.

Verifying

node --check idleon-fishing.user.js
python3 tools/suite/build.py        # idempotent; diff should be empty

Against the offline module: 1024/1024 positions agree exactly on the catch
window across four tolerances; refitAim recovers a synthetic lane (u0 20, uW
310) to three decimals; the lead learner records +6 rungs for a release six
rungs late, −6 for six early, and drops both unattributable and ambiguous casts.
The shipped tracker, fed fish positions from the offline module frame by frame
and scored against the game's own bob, calls casts that catch 97.6% of the
time at both 50 and 30 fps. A 5% slow page clock costs 93.3%, a 20% slow one
82.5% — still well clear of 53%, so lag degrades this rather than inverting it.

Replayed over the fishing recording, 408 frames at 6 fps: 348 with fish, 347
with a plan, tracking live on 144, up to 4 tracks at once, every plan a single
run, no exceptions.