Cardano Updates

One bisect-safe slice extends Transaction.Ledger with a
ScriptWitnesses carrier (native inputs, staking script, mint
sources, output reference script), plumbs it into the ledger
body and witness-set lenses listed in data-model.md, and proves
body+wits byte-equality against the legacy mkUnsignedTx route
on the six representative scenarios from spec.md. mkUnsignedTx
and mkUnsignedTransaction stay byte-identical to origin/master;
lib/integration/** is not touched; Dijkstra remains pendingWith.

research.md records the body-field map, the witness-set lens
choice (witsTxL . scriptTxWitsL at unsigned-build time), the
single-record API extension, and why the live-boundary
diagnostic returns "none" for this slice (pure-CBOR comparison
of two builders, no system boundary to cross).

Smallest first slice toward unblocking issue #5285 Story 1: plumb a
mint set through buildLedgerTx and buildLedgerTxRaw, translate the
wallet's (TokenMap, TokenMap) mint/burn pair to a single signed
ledger MultiAsset, and feed the real value from mkTransactionLedger
and constructUnsignedTxLedger. Script witnesses remain out of scope.

Artifacts: spec.md (2 P1 stories, 9 FRs, 5 SCs), plan.md (113 lines,
constitution pass, 6 design decisions), research.md (Phase 0 from
Explore agent: signed/unsigned asymmetry between TokenMap and
MultiAsset is the load-bearing surprise), data-model.md (netting
rule for mint/burn overlap), contracts/ledger-mint-translation.md
(P1-P8 converter properties, B1-B3 builder properties), tasks.md
(15 tasks across 5 phases mapping to three bisect-safe commits),
requirements checklist, quickstart for reviewers.

Pre-flight grep on the implementation pass showed lib/wallet/src/Cardano/
Wallet.hs already imports certificateFrom{Delegation,Voting}ActionLedger
from Cardano.Wallet.Shelley.Transaction.Ledger and its 4 callsites at
2729, 2738, 3533, 3542 already use the *Ledger variants. The Phase-0
research had it wrong; plan.md, quickstart.md, and tasks.md inherited
the error. Spec.md was correct (line 18 already named the single
caller).

Slice scope shrinks from 'edit two files + delete two + cabal prune' to
'edit Shelley/Transaction.hs + delete two + cabal prune'. SC-001/SC-002,
FR-003, and the byte-equivalence proof are unaffected.

Phase 0 + Phase 1 speckit artifacts for issue #5285.

- plan.md (109 lines): three vertical slices, constitution check clean,
  Story 1 unblocked, Stories 2-3 blocked on Transaction.Ledger
  prerequisite ACs in #5243.
- research.md: full cardano-api symbol inventory in
  Shelley/Transaction.hs (~40 symbols across 4 imports), per-Story
  grouping, ledger replacements located.
- data-model.md: cardano-api -> ledger type map per Story.
- contracts/module-interface-invariants.md: exports that MUST stay
  signature-stable across all three slices.
- quickstart.md: ordered implementation walkthrough for Story 1
  (the unblocked slice; 5 callsites in Shelley/Transaction.hs,
  4 in Cardano/Wallet.hs, plus cabal prune and helper deletion).

## Summary

- use a short unique `/tmp/cwbench.XXXXXX` root for the Linux Memory
Benchmark before invoking the existing benchmark command
- keep non-memory benchmark matrix entries on `${{ runner.temp }}`
- ignore local `llm/` review artifacts and remove the tracked
`llm/reviews/5279/gate.sh` file inherited from the rebased base

## Verification

- `actionlint .github/workflows/linux-benchmarks.yml` with custom
runner-label warnings ignored
- old representative socket path: 120 bytes; new representative socket
path: 94 bytes
- `/code/llm-settings/shared/skills/pr/references/commit-gate.sh HEAD`
- Linux Benchmarks workflow dispatch passed end to end before the
rebase:
https://github.com/cardano-foundation/cardano-wallet/actions/runs/25724369765
- Memory Benchmark artifact contains `memory-bench-results.csv`,
`memory.hp`, `memory.svg`, `memory.log`, and `memory.error.log`

Fixes #5283

Use a short mktemp directory under /tmp for the Linux Memory Benchmark before running the existing benchmark command. This avoids the cardano-node Unix socket path limit while keeping other benchmark jobs on runner.temp.

Ignore local llm review artifacts so they do not enter the repository.

Verification: actionlint accepted linux-benchmarks.yml with repository custom runner-label warnings ignored. The representative issue path is 120 bytes, and the new representative socket path is 94 bytes.

Live verification: Linux Benchmarks workflow dispatch 25724369765 passed end to end and uploaded memory-bench-results.csv, memory.hp, memory.svg, memory.log, and memory.error.log.

Fixes #5283