Fix #67: context export named parent uuids that are not in the network

[adamjohnwright]

Closes #67.

append_regulators decomposes each catalyst and regulator to its terminal
members and wires the edges from those member uuids. The context export
wrote the uuid carried on the raw fetch row instead — the undecomposed
parent — so it named nodes that do not exist in the network.

Measured on a ten-pathway catalog before the fix:

role	rows	orphaned
input	3,641	0	0.0%
output	3,491	0	0.0%
catalyst	2,228	2,228	100.0%
regulator	1,018	1,018	100.0%
TOTAL	10,378	3,246	31.3%

Every consumer asking "which node is this entity, at this reaction, in this
role" got a wrong answer for exactly the two roles that carry the causality.

The fix

Catalyst and regulator rows are now read from the emitted network rather than
from the fetch rows, which makes the export correct by construction: it
reports what was wired, not what was requested. Those edges carry no
edge_reaction_id, but their target is the virtual-reaction node, so the
reaction is recovered through reaction_id_map exactly as the input/output
rows already do.

Two guards, because the quiet failure is what let this survive:

• the export refuses to write when a row names a node absent from the
  network, and reports the count by role rather than shipping it;
• omitting the network while catalyst rows exist now raises, instead of
  falling back to the fetch rows and silently restoring the bug.

The bug was held in place by a test

test_export_node_reaction_context asserted that the parent uuid c1
appears in the export — the exact wrong behaviour — so the suite was green
while a third of the export was broken. That test is inverted and kept as the
record, alongside a new one that fails when the network and the export
disagree.

Verified as a negative control: reintroducing the parent uuid makes the new
test fail; restoring the fix makes it pass.

Verification

• 167 tests pass in the no-database tier.
• Regenerated the ten-pathway catalog: orphans 3,246 → 0, and coverage
  0 → 790 of 790 distinct (node, reaction, role) triples. Zero orphans
  is trivially achievable by dropping rows, so coverage was checked
  separately.
• The row count falls from 2,228 catalyst rows to 571 and that is not loss:
  3,758 catalyst edges collapse to 790 distinct triples because the same
  member catalyses the same reaction through many parallel virtual-reaction
  instances. The old count enumerated parent × reaction, a different and
  wrong population.
• Export-only: all ten regenerated networks are structurally identical to
  the originals — edge counts, node counts and the
  edge_type/sign/and_or/stoichiometry signature match 10 of 10 — so no
  benchmark number moves because of this.

Known, not fixed here

Catalyst and regulator edges carry no edge_reaction_id at all (3,758 and
1,124 of them). The export now routes around it via the target reaction node,
so nothing is blocked, but the column is empty where it should not be.

🤖 Generated with Claude Code