Match Vode statuses and rule keys exactly, not by prefix

[cemde]

Bug

Two regexes in _vode.py were anchored only at their start, so both matched by prefix where they meant to match the whole string.

Ruleset.filter used re.match(pattern, status). The built-in ruleset registers its forward-init rules under the pattern "init", so re.match accepted any status beginning with init.

Vode.set built its rule pattern as f"...<-\s*({key}.*)". The trailing .* means the key group matched any rule whose right-hand side merely starts with the key.

Impact

A user who names a phase initialise or init_weights silently gets the h, u <- u rule, which copies the prediction into the node value. The prediction error, and therefore the energy, is identically zero for that whole phase.

For keys the damage is doubled. A node with two incoming activations, u from the layer above and u2 from a skip connection, has u2's rule fire with u's value. And because a rule matched, the len(rules) == 0 fallback never runs, so u is never stored under its own name at all. With ruleset ("z <- u2",), set("u", 7.0) left cache["z"] == 7.0 and cache["u"] unset.

Fix

Both matches become re.fullmatch, and the key group gains an explicit optional tail for the :transformation suffix:

if re.fullmatch(_pattern, status) is None:      # was re.match
...
if _match := re.fullmatch(rule_pattern, _rule):  # was re.match
...
_rule_pattern = f"(.*(?<!\\s))\\s*<-\\s*({key}(?::.*)?)"   # was ({key}.*)

fullmatch restores the documented meaning of the patterns as regexes: .* is the way to say "any status", which was meaningless while every pattern carried an implicit one. (?::.*)? is non-capturing, so _match.group(1, 2) still yields the same targets and transformation.

Why not a trailing $

$ also matches immediately before a trailing newline, so it is strictly weaker than fullmatch. It would also leave Ruleset.filter applying whole-string matching to its status argument and prefix matching to its rule argument, with the anchoring obligation pushed into an undocumented caller convention that Vode.get does not follow.

Closes #69

[liukidar]

LGTM