feat(meta-controller): Phase 7 — Pareto + Bayesian routing across (prover × coprocessor × aspect)

Closes the architectural loop: 11 coprocessor backends + 100+ provers
existed before this commit but the routing brain only knew about Math
and 3 aspects.  Phase 7 wires:

**Full coprocessor registry** — `MetaController::new()` iterates the
known `CoprocessorKind` variants and asks `CoprocessorFactory::native`
for each, registering whatever's available.  10 always-on (Math /
Vector / Tensor / Crypto / Physics / Dsp / Io / Graphics / Audio / Fpga)
+ FlintMath when `--features flint`.

**Broad aspect taxonomy** — 22 aspect → coprocessor-op mappings cover:
- Arithmetic (gcd/lcm/modexp/modinv/primality/factor)
- Polynomial algebra (FLINT)
- Linear algebra (Vector dot/norm, Tensor matmul/det)
- Crypto (sha256/blake3/ed25519-verify)
- Physics (RK4 step, harmonic energy)
- Signal (FFT, Hann window)
- IO (file hash, line count)
- Visualisation (proof-graph SVG)
- Audio (completion chime)
- Hardware (yosys synth)

FLINT-routed aspects gracefully filter to empty when the backend isn't
linked — no failure path; controller routes around.

**Pareto + Bayesian** — `plan_with_pareto(goal, candidates, default_timeout,
default_trust)` builds a `ProofCandidate` per prover, populates predicted
proof_time from the StatisticsTracker mean (or default when absent),
runs `ParetoFrontier::compute`, picks from frontier via weighted_rank
favouring time.  Falls back to `plan` when no candidates supplied.
Reuses the existing pareto.rs + statistics.rs surfaces — no new infra
invented.

**Outcome recording** — `record_outcome(plan, goal, outcome, elapsed_ms)`
flows Success/Timeout/Failure back to the StatisticsTracker keyed by
(prover, primary-aspect-as-domain).  No-op when stats absent.

**Plan enrichment** — `Plan` now carries:
- `estimated_timeout_ms` (Bayesian per-tuple, falls back to default)
- `rationale` (human-readable diagnostic — pareto vs heuristic vs preferred)

Tests (all green):
- registers_all_native_coprocessors (10 or 11 depending on feature)
- aspect_routing_covers_each_kind_at_least_once (10 spot-checks)
- plan_with_pareto_picks_from_candidates
- plan_with_pareto_falls_back_when_no_candidates
- record_outcome_with_no_stats_is_noop
- record_outcome_updates_stats_when_present (success rate 2/3 verified)
- estimate_timeout_uses_stats_when_present (Bayesian estimate < 5 s after
  ten 100 ms successes; was hitting 30 s default before stats)
- flint_aspect_hint_filtered_when_backend_absent

panic-attack assail clean (0 weak points).  Pre-existing prover-side
trust pipeline + 700+ lib tests unchanged.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: hyperpolymath

.machine_readable/6a2/STATE.a2ml

@@ -87,9 +87,10 @@ notes = "5 new CoprocessorKind variants (Dsp/Io/Graphics/Audio/Fpga) with full R
 [[expansion-roadmap.phase]]
 id = "phase-7"
 title = "Meta-control — Pareto over (prover × coprocessor × aspect), Bayesian timeout per-tuple"
-status = "blocked-on-phases-1-6"
+status = "complete"
+landed = "2026-04-27"
 adds-backends = []
-notes = "Extension to existing MetaController Plan + run_preconditions seam."
+notes = "MetaController extended: registers all 11 native coprocessors at construction; aspect routing taxonomy expanded from 3 entries to 22 covering arithmetic / polynomial-algebra / linalg / crypto / physics / signal / IO / visualisation / audio / hardware. Reuses existing `verification::pareto::ParetoFrontier` and `verification::statistics::StatisticsTracker` (no new infra invented). New API: `MetaController::with_stats(tracker)` plugs in a shared StatisticsTracker; `plan_with_pareto(goal, candidate_provers, default_timeout, default_trust)` ranks candidates by predicted (proof_time, trust, memory, steps) and picks from the Pareto frontier; `record_outcome(plan, goal, outcome, elapsed_ms)` flows results back to the tracker. `Plan` now carries `estimated_timeout_ms` (Bayesian per-tuple) and a `rationale` string for trust-pipeline diagnostics. FlintMath aspect-hints filtered when backend absent (graceful no-feature fallback). 8 new tests, all green; full lib still compiles cleanly."
 
 # ═══════════════════════════════════════════════════════════════════════════════
 # CRG (Critical Reception Grade) advancement — D→C→B