`random.Random` has no per-instance RNG, and the state machinery already exists

Row 7 of devdocs/dev/tsp-compile-wall-inventory-2026-09-20.md, second half. The FIRST half — a silent wrong value — is fixed and is not this ticket; see devdocs/dev/tsp-rows-4-8-what-shares-a-cause.md. In short: the intercept table's callers lower-cased the member, random.Random folded onto random.random, and random.Random() compiled to a float. The member is now matched as spelled, so the call is refused rather than silently wrong.

What the callers want

tsp/smoke.py:60        self.rng = random.Random(seed)
tsp/commentary.py:64   rng = random.Random(mission_id)

Both want the same thing and it is the reason Random exists in CPython: a stream that is reproducible from a seed and independent of the global one, so a seeded replay is not perturbed by anything else in the program that draws a number.

random.seed(n) plus the module-level functions is NOT a substitute, and offering it as one would be the wrong fix: it makes the replay depend on every other consumer of the global stream.

The state is already here, in full

lib/rtl/random.pas already declares a per-instance state record and its whole API:

TRandomState
RandomStateSeed / RandomStateRandomize / RandomStateNext
RandomStateRange / RandomStateRange64 / RandomStateDouble
RandomStateBytes / RandomStateSplit

So this is not "implement an RNG". It is a class wrapping a record that exists, with the Python method names (random, randint, randrange, uniform, choice, shuffle, seed) on it.

THE ACTUAL BLOCKER IS REACHABILITY, NOT THE CLASS

Plain import random is consumed-only (PyImportRootPlainIsConsumedOnly), so random is not bound to lib/rtl/random as a unit alias, and a qualified random.Random has no unit to find a class in. That is consistent with the diagnostic the fixed compiler now gives — undefined variable (random), i.e. the name is not bound at all rather than bound to something without that member.

Note the asymmetry already recorded in pyparser.inc: the FROM-spelling (from random import randint) takes the unit arm, while the PLAIN spelling is consumed. So the two import spellings of the same module reach different machinery, and a fix has to decide which one random.Random goes through. Whoever takes this should read that comment before choosing — the two spellings have been out of step in both directions before (bug-n-from-sys-import-fails-while-import-sys-works).

Two name hazards, both real, both cheap to hit

Random is a Pascal RTL function name. A class of that name inside unit random collides with System.Random.
A class named after a unit in its own uses clause silently miscompiles its own constructor into a self-call — lib/rtl/pil.pas records the measurement and works around it with Image = TPILImage. A class called Random in a unit called random is that shape.

The established workaround is the alias, which carries its own known cost: type(x).__name__ then reports the Pascal spelling (bug-n-a-shim-class-reports-its-pascal-spelling-from-type-name). Acceptable here — neither caller introspects the RNG — but it should be a decision and not a discovery.

What would retire this

smoke.py past line 60, and a fixture asserting the property the class exists for: two Random(1234) instances produce the SAME sequence, and drawing from the global random in between does not change it. A test that only checks "returns a number in range" passes against the global stream and certifies the bug. Seed with a value whose first draw differs from the global stream's, or the row cannot fail.

ABSORBED FROM ITS DUPLICATE, 2026-09-21

feature-n-random-random-the-per-instance-rng-class-is-absent (p40) was the same feature filed a day earlier by the same seat, and is resolved as superseded by this one. Both were mine; neither carried a supersedes edge, and both sat in BOARD.md as live work. Reported by frankz-e5 off the TSP survey.

Why THIS one is the survivor rather than the older: the p40 summary framed the gap around a CONDITION that no longer holds — that the NilPy member lookup folds case, so Random matched the module-level random() and the program was refused with pyrandom_random takes fewer arguments than were given. That fold was fixed on 2026-09-21 (the member is now matched as spelled), so a reader taking p40 at face value would have gone looking for an arity message that the compiler no longer emits. Its diagnosis was correct when written and its repro still reproduces — with a different, better error.

The one thing in p40 worth keeping, because it is a CLASS and not this ticket: any absent member whose name differs only in case from a present one used to be reported as an ARITY error about a call the source never wrote, which sends the reader to the call site instead of to the missing member. That is retired as a live hazard by the as-spelled fix, and it is recorded where a retired hazard belongs — the logbook and the case-sensitivity work — rather than in a feature ticket, so it cannot be mistaken for something still to do.

p40's repro, kept because it is three lines and still the shortest statement of what is wanted:

import random
r = random.Random(7)
print(type(r).__name__)      # CPython: Random