fix(autodiff): close four gaps between the two expansion passes

[AlekseiChirkovVention]

Close four defects that only became visible once the whole change was read as one artifact.

Why these were not caught earlier

Each branch in this series was reviewed against its own diff. That bounded review is what kept
each one tractable, but it means no reviewer was positioned to see an inconsistency between
two branches. All four fixes here are of that kind, and each was reproduced against the code
before being approved — none is speculative.

The four

The two passes held contradictory contracts for the same invariant. The gradient path
indexes value ids a program merely reads and documents why; the forward path indexed only
values it produces. So a generated identifier could collide with a value a node already reads,
and the result was not an error — the consumer was silently rewired to read a generated
constant instead. Both passes now derive collision candidates from one shared helper covering
every identifier an artifact mentions: produced, read, declared as an input, declared as an
output. The gradient side gained a position it was also missing.

Sharing one helper is the point. Correcting the forward side by copying the gradient side's
logic into a second place would have left the same drift to recur.

The constant reader was not authoritative about its own node's declared type. A node
declaring one dtype and shape while its parameters said another read back as the parameters,
with no complaint — though the declaration is documented as the single authority a consumer
materializes from. A handler would have allocated one tensor while the graph declared another.
The cross-check runs after parameter validation, so every existing rejection keeps its category
and a node with no declaration still reads.

An oversized declared shape raised a bare builtin. An operand declaring very large
dimensions passes every clause of the supported-reduction predicate and then overflowed inside
the reciprocal. It now fails validation naming the node and the count.

The threshold is deliberate: counts are rejected only when they cannot be converted, not when
they exceed the exactness limit. A count above the latter still expands to well within the
documented tolerance, so rejecting there would have narrowed the supported domain rather than
guarded it — a behaviour change wearing a guard's clothing.

A documented requirement had no test. The documentation calls out that the rank-reducing
tier needs a reshape handler in both artifacts, because the derivative transform's own seed
reshape survives expansion. That behaviour was correct and pinned by nothing; the missing
fail-closed case now covers it.

Why this is safe on top of eight reviewed branches

These are guards. For any well-formed artifact both passes emit exactly what they emitted
before, and every pre-existing test passes unmodified — that is the regression signal, and
review confirmed it by expanding sixteen well-formed artifacts through all four entry points
against the previous implementation and finding zero differences.

Focused suite 895 → 910, exactly the new cases. Broad non-integration suite: 1161 passed,
9 failed, 1 skipped — the same 9 environment-dependent failures, identical in composition, that
this series has carried from its first branch.

Two test amendments, both disclosed

One pre-existing test was strengthened, as this fix-task's own plan required: the agreement test
built both sides from a single dict and so could not fail. It now spells them independently and
asserts a disagreeing declaration is rejected. The three original assertions survive verbatim.

One test authored earlier on this branch had its fixture corrected: it fed the reserved
identifier into the very region being expanded, which stopped the region matching, so nothing
was minted and the collision it asserts could never fire. It was red, but for the wrong reason,
and would have stayed red permanently. Assertions unchanged. Worth stating plainly, because a
self-authored test that quietly stops testing what it claims is the same class of defect this
branch exists to close.

Stack position

Stack position 9 of 9 — based on chore/mean-expansion-regression-coverage.

The tip of the series. Cannot merge into main before the eight branches below it.