ci: reduce duplicated build and restore work without changing coverage #783

Context

PR #782 highlighted that the current workflows are functionally healthy, but still perform some equivalent setup/build work in isolated jobs. We should reduce runner time conservatively, without skipping checks, changing required check names, weakening test coverage, or introducing path-based exclusions.

Baseline from the PR #782 runs:

• Build candidate images: 2m11s
• ARM64 identity directory smoke: 2m11s
• Build & Test: 2m34s
• Format check: 1m39s
• Frontend lint (smoke): 33s
• Frontend a11y: 40s

Runs:

• CI: https://github.com/pedrosakuma/B3TradingPlatform/actions/runs/31196291542
• Docker: https://github.com/pedrosakuma/B3TradingPlatform/actions/runs/31196291503

Proposed conservative scope

1. Reuse the trading-host Buildx cache in the ARM64 smoke build

The PR candidate job already builds backend/src/B3.Trading.Host/Dockerfile for linux/amd64,linux/arm64 using the trading-host GHA cache scope. The dependent identity-directory-arm64-smoke job builds the same Dockerfile for linux/arm64 again, but reads/writes a separate trading-host-arm64-smoke scope.

Allow the ARM64 smoke build to consume the completed trading-host cache while preserving its current independently loaded image and runtime smoke assertions. It may keep a separate write scope if that avoids cache contention.

2. Add the existing NuGet package cache pattern to the format job

build-test restores with an explicit ~/.nuget/packages cache. The independent format job performs another full restore without that cache configuration. Reuse the same cache-key convention in format; keep dotnet format --verify-no-changes unchanged.

3. Benchmark frontend bundle reuse before changing job dependencies

The frontend smoke and accessibility jobs independently install the root frontend dependencies and build frontend/dist. Reusing the bundle through an artifact could reduce runner work, but adding a producer/consumer dependency may increase wall-clock critical-path time because these jobs currently run in parallel.

Prototype or measure this separately and only adopt it if:

• total runner work decreases;
• end-to-end PR latency does not materially regress;
• the a11y job still audits the exact bundle produced from the PR commit;
• frontend tests, container smoke, axe and Lighthouse coverage remain unchanged.

Non-goals

• no path filters or conditional skipping of required workflows;
• no removal or consolidation of functional checks merely to save time;
• no reduction in architectures, conformance profiles, tests, coverage collection, axe, or Lighthouse;
• no changes to product/runtime behavior;
• no broad workflow rewrite.

Acceptance criteria

• ARM64 identity smoke consumes reusable layers from the preceding trading-host candidate build while still loading and booting a real linux/arm64 image.
• The format job uses the repository's existing NuGet cache convention and still performs its own restore plus --verify-no-changes.
• Required check names and branch-protection behavior remain unchanged.
• All current CI, Docker, conformance, ARM64 runtime, frontend, and accessibility checks remain enabled.
• Before/after timings are recorded from at least two representative PR runs.
• The change demonstrates lower runner time and does not materially increase the PR critical path.
• Any frontend artifact reuse is omitted if measurements show a wall-clock regression.