feat(skills): Add dependency-redundancy analysis to the user plugin

[lyfne123]

Summary

The user plugin can now explain where a run's time went, but not whether the
graph it executed carried orderings nothing needed. dependency-redundancy
adds that audit over deps.json: which edges a longer path already implies,
which of those are safe to drop, and — the part the tool does not say on its
own — when its answer is uninformative and must not be reported as "already
minimal".

Changes

• plugins/pypto-user/skills/dependency-redundancy/: the skill. It orders the
  work so uninformative cases are excluded before a mode is chosen — read the
  edge-source mix and the graph depth first, then select — and covers the four
  edge modes, the validation gates, and an interpretation table naming the
  action each outcome licenses.
• plugins/pypto-user/lib/dfx/capture.md: note that a structural analysis
  consumes deps.json alone, so the two analyses sharing that file do not both
  have to capture timing.
• plugins/pypto-user/.codex-plugin/plugin.json: the long description covers
  the fifth workflow. No version bump — that belongs to the release workflow.
• tests/: register the skill in EXPECTED_USER_SKILLS.

No behavior change to any existing skill; the new skill is additive.

Why the ordering matters

Census over 64 dependency graphs captured with --enable-dep-gen across
tests/st/a2a3/{tensormap_and_ringbuffer,host_build_graph} on real silicon:

	edges	structurally redundant	reduced reports	reduced_dataflow reports
all 64 graphs	43784	5839	722 (12%)	2709 (46%)
batch paged attention	6128	2032	0	992
paged attention unroll	1280	256	0	0
dep-gen chain (n=391)	783	391	391	391
64×64 explicit fanin	4096	0	0	0

The mechanism works — the chain cases prove it finds and prints real redundant
edges. What it cannot do is say when its own answer is uninformative. Creator
protection is per-pair, and Step-A creator retention emits a creator row for
exactly the pairs a dataflow dependency also covers, so a mixed-source graph is
no safer than an all-creator one: every one of batch paged attention's 2032
redundant pairs is creator-annotated, and the documented default mode reports
removed 0 on a graph where 992 edges are provably droppable.

18 of the 29 graphs that contain redundancy report 0 from every mode, because
their redundant pairs have INPUT consumers while the byte-level proof only
licenses INOUT. Those edges are correctly retained — they hold a creator
alive — but nothing in the output distinguishes that from an already-minimal
graph. The skill turns this into a rule: never report from reduced alone, and
when both modes return zero, say "nothing removable" rather than "nothing
redundant".

Verification

• python -m unittest discover -s tests: 232 tests, OK. Run inside the push
  transaction at each pushed commit with ruff check tests,
  ruff format --check tests, and pyright — all clean, 0 errors.
• On-device corpus. The census above: the full a2a3 scene-test suite run
  under the shared device queue with --enable-dep-gen, then every graph
  audited by following the skill literally. The per-mode counts come from the
  shipped CLI; the structural column comes from running the same reduction with
  the creator protection bypassed.
• Mode semantics on a minimal graph. A three-task chain with a direct
  1 -> 3 edge: reduced and omitted both identify and print 1 -> 3;
  flipping that edge's source to creator makes both report 0, reproducing the
  trap at a scale small enough to read.
• Cycle handling. A cyclic graph warns on standard error, emits the full
  graph, and exits 0 — which is why the skill tells the reader not to take the
  exit status as proof a reduction ran.
• Output naming confirmed: a reduction writes deps_viewer_<mode>.txt rather
  than overwriting a full-graph render beside it.