Add exact persistent representation broker routes

[Inserian]

QSA exact-representation progression remains on this draft/unmerged staging PR.

Accepted authorities:

• factor broker: 423be022f348f102d700dbcc368a7545744419f6
• direct persistent TTN: f723ea81bd0374f13eca4869118f195550c08c32
• TTN marginal resource certificate: 30b0cbbc4617058947ea9b6099c8881068a4fe6b
• marginal-only TTN execution broker: 487858b1baea9eb9604a79fc973409f05eb53f33

Accepted factor broker order remains:
ChainTransfer -> AffineXOR -> DecisionDiagram -> VariableElimination

Public ExactFactorBrokerRoute values remain stable: ChainTransfer=0, DecisionDiagram=1, VariableElimination=2, AffineXOR=3.

Direct TTN — accepted direct representation

TreeTensorState is accepted as a persistent exact state representation for branching entanglement.

Accepted TTN contract:

• deterministic balanced binary qubit tree;
• canonical QSA OperationCode dispatch for X/Y/Z/H/S/Sdg/T/Tdg/Rx/Ry/Rz and exact CNOT/CZ;
• arbitrary-leaf CNOT/CZ through exact operator-Schmidt rank-2 branch insertion along the unique tree path;
• touched path bonds double exactly and internal path tensors enforce branch equality;
• no SVD, truncation or approximation;
• explicit max-bond, state-scalar, materialization and marginal-workspace caps;
• controlled-gate resource growth is preflighted before allocation or mutation;
• failed resource gates preserve accepted state/statistics;
• exact amplitudes/probabilities and selected-qubit marginals by double-layer tree contraction;
• workspace generation and per-node shapes are revalidated, including cross-state reuse.

All 17 exact-head workflows were green on direct authority f723ea81bd0374f13eca4869118f195550c08c32.

Decision-bearing direct TTN evidence: run 31642042492, artifact 9159238027, ZIP digest sha256:aedbebf6e9c10926c883cd85b610016e6632585fe25656866e5e362cf7664225, embedded evidence SHA-256 7441c9cda7611222a0b7ddcfd9acad5f5cbb03ca2aea20c5795969cdc9f71630.

Accepted direct evidence:

• matched 16-qubit carrier: 12 controlled gates, max path 4, max bond 8, 867 state scalars; TTN setup 0.012339 ms vs dense setup 2.452984 ms; prepared TTN query 0.02588327 ms; plan 18,120 bytes; workspace 14,352 bytes; amplitude error 2.78e-17; marginal error 0;
• 4,096-qubit carrier: 1,024 independent four-leaf branching blocks, 3,072 exact controlled gates, max path 4, max bond 8, 222,207 state scalars; setup 2.739705 ms; selected-block marginal query 7.6754504 ms; plan 4,636,680 bytes; workspace 3,674,112 bytes; selected probability 0.0119166075286; independent block-control error 2.75e-15.

QRegister amplitude comparisons use QSA's established physical-state equivalence up to one global phase; marginal probabilities remain strict. TTN is exact and topology/workload-specific. Repeated overlapping paths can grow bonds exponentially and must fail closed. No universal or quantum-advantage claim is made.

TTN marginal resource certificate — accepted

TreeTensorMarginalPlan is accepted as a bounded preconstruction resource certificate and prepared marginal wrapper. It builds bounded O(n) topology/resource vectors but does not allocate TTN amplitude/tensor storage during certification; it is not described as allocation-free.

The certificate predicts and verifies the accepted executor's qubit/node counts, state scalar count, maximum bond dimension, controlled-gate count, maximum controlled path length, generation count and double-layer marginal-workspace scalar requirement. It fails closed on any prediction/executor disagreement.

Focused certificate run 31643019804 is green on Ubuntu 24.04, Windows 2022, macOS full-core and ASan/UBSan. All 18 exact-head workflows were green on 30b0cbbc4617058947ea9b6099c8881068a4fe6b.

The focused gate includes the independent workspace boundary: in the repeated 4-qubit cross-root carrier, TTN state storage is exactly 228 scalars while marginal workspace requires 259; cap 259 succeeds, cap 258 still permits direct state construction but direct marginal rejects, while the certificate rejects before TTN construction/query.

Marginal-only TTN execution broker — accepted

ExactPreparedProbabilityPlan now appends public execution route TreeTensor=8 without renumbering existing routes: Register=0, CausalPauli=1, TensorNetwork=2, PersistentMPS=3, PhaseGraph=4, UniformMagnitude=5, BasisPermutation=6, Stabilizer=7.

Promotion scope is deliberately narrow:

• TTN is reachable only from marginal preparation / marginal_probability_from_zero;
• existing BasisPermutation -> UniformMagnitude -> Stabilizer -> PersistentMPS marginal authority remains ahead of TTN;
• TTN is attempted only after MPS rejection and before dense QRegister fallback;
• full-basis probability preparation cannot select TTN;
• TTN rejection remains visible and falls through to exact Register;
• unsupported trajectory operations still prove the Register fallback path;
• direct TTN and certificate implementations remain unchanged from their accepted authorities.

The production promotion itself is compact: include/qubit/qbroker.hpp adds 4 lines and src/qbroker.cpp adds 22 lines, with zero production deletions relative to the accepted certificate authority. The remainder is focused tests and evidence carriers.

All 20 exact-head workflows are green on 487858b1baea9eb9604a79fc973409f05eb53f33, including Build/Test, full Runtime Evidence, inherited Exact Execution Broker, TTN direct/QRegister/certificate gates, Tensor, factor routes, Causal, Pauli, Structural, Dense COW and Brain Recovery. The focused marginal-TTN broker gate is green on Ubuntu 24.04, Windows 2022, macOS and ASan/UBSan.

Decision-bearing broker evidence: run 31645271142, artifact 9160517879, ZIP digest sha256:c3830b40182be055f0a0636b5c2110c0094a7803d7edc536075e527f510d0d9a, embedded evidence SHA-256 642bcfb882e605811b4b8bcc252d9415f5e9553fa6d183db7f20367d0ac2d63f. GitHub metadata, the independently downloaded ZIP and the embedded receipt were reconciled.

Accepted broker evidence:

• matched 16-qubit / 20-operation carrier: route TreeTensor; broker setup 0.011996 ms vs direct TTN 0.010943 ms; broker query 0.02077274 ms vs direct TTN 0.01973342 ms; broker plan 19,611 bytes vs direct 18,271; error 6.94e-18;
• 4,096-qubit / 5,120-operation branching carrier: route TreeTensor; broker setup 2.993234 ms vs direct TTN 2.530621 ms; broker query 6.7874135 ms vs direct TTN 5.7179153 ms; broker plan 4,638,171 bytes vs direct 4,636,831; selected/control probability 0.0119166075286; error 2.75e-15.

The measured routing wrapper is not free: about 9.62% setup / 5.27% query overhead on the 16-qubit carrier and 18.28% setup / 18.70% query overhead on the 4,096-qubit carrier, with a fixed 1,340-byte retained-plan delta in these measurements. Direct TTN remains the stronger path when the caller already knows the representation. Broker value is exact eligibility/route selection and fail-closed fallback, not a claim that indirection is faster.

Next bounded phase — direct exact QTT functional representation

The next new representation frontier is QTT/tensor-train scoring for exponentially indexed but low-rank structured functions. Keep this phase direct-only first; do not add it to ExactExecutionBroker until independent evidence earns promotion.

Initial contract:

• exact binary quantization of a logical 2^n index into tensor-train cores;
• construct only from inputs carrying an exact low-rank certificate or from exact generators whose TT ranks are known by construction;
• no approximate TT-SVD, truncation, tolerance-based rank guessing or silent compression;
• explicit max-rank / max-core-scalar caps and fail-closed rank-growth checks;
• exact point evaluation, norm/partition-style contraction and a bounded marginal/reduction only where the rank certificate supports it;
• strong controls against direct closed-form evaluation, current MPS/Tensor where the workload overlaps, and dense materialization only at small sizes;
• include no-benefit/adversarial high-rank inputs where QTT must reject rather than densify.

TreeTransfer remains rejected/parked. Prior accepted Tensor, ChainTransfer, DecisionDiagram, AffineXOR, TTN direct/certificate/broker and cost-aware tensor-planning mechanisms remain closed unless new evidence exposes a distinct mechanism failure.

Keep PR #62 draft and unmerged. Do not merge without explicit authorization.