Write two explainers through the passes that had never run

[abernier]

Closes #214. Part of #192, on top of #213.

--explain has always specified that it does not read the source and write: it
runs teach in sub-agents first, and writes only from what they distil. Nothing
had ever gone through it. aquarium's explainer was written by direct reading
and kept on purpose — so the orchestration was untested at the point where it
was about to be used 170 times.

The two examples

Chosen to be awkward rather than photogenic, one on each side of #214's split:

• inverted-stencil-buffer — the technique is already understood, and it is
  a direct sibling of aquarium's stencil mask, so the output can be read
  against a hand-written explainer on the same technique. This is the control.
• gpgpu-curl-noise-dof — genuinely opaque: an off-screen float target, two
  hand-written shader materials, and a noise implementation nobody reads twice.

Both cleared in one pass. The cap of three never engaged.

What the passes returned that a direct reading would have gotten wrong

• there is no ping-pong in gpgpu-curl-noise-dof. One target, no feedback,
  the seed re-read and never overwritten. It is a position field, not a particle
  simulation — and the directory name says the opposite. This is the single most
  likely sentence in a directly-read explainer, and it would have been false.
• its curl noise is normalised to unit length, which destroys the
  divergence-free property curl noise is cited for.
• its fov is not a field of view; it gates visibility and saves no
  simulation cost.
• Mask's three stencil ops are all Replace, so the stamp is unconditional and
  ids overwrite rather than layer.
• inversion belongs to the masked material, not to the mask. drei's own
  documentation describes it the other way round.

The last two come from the control — the example where direct reading was
expected to be sufficient, because that is the bet aquarium made.

The decision

The remaining 168 go through the orchestration. The reasoning, the cost and
the three reservations are in the comment on #214. The second commit here
carries the four corrections the runs found, so explain.md describes what
actually happens rather than what was imagined.

Notes

• Each example's description and apis are written from the same distillation
  as its explainer, per the one-reading rule. Neither was in UNDESCRIBED, so
  both lines are rewrites of weak ones and the list stays at 33.
• lint:metadata covers the two new CONTEXT.md files. A planted identifier
  fails it — the guard was confirmed rather than assumed.
• explain.md line 12 still says an explainer is "not a programme across 170",
  which Distilled docs: one line to choose, an explainer once chosen #192's plan now contradicts. Left alone deliberately: that is the scale
  question, and it belongs to its own issue.

🤖 Generated with Claude Code

https://claude.ai/code/session_01YQmbpRu6oCNrLiUQvhz5x6