Add ranked analyze map evidence bundles (--rank-probes, --evidence-bundle)

[yuechen-li-dev]

Motivation

• CTC-01 and dogfooding showed coarse six-view maps are useful but global higher-resolution grids are too slow for tight LLM loops, so the tool must prioritize local, ranked probes instead of brute-force refinement.
• The goal is to present LLMs with compact, actionable measurement evidence (ranked probes, suggested sections, local map windows) so reasoning can focus on hypothesis testing rather than parsing huge grids.

Description

• Add CLI flags --rank-probes and --evidence-bundle (wired through CliRunner), with --evidence-bundle implying ranked output and --rank-probes enabling scoring.
• Extend analysis models (StepAnalysisModels.cs) and StepAnalyzer to emit RankedProbes, an evidenceBundle structure, and a compact point-probe summary (pointSummary) while preserving full hits and existing detailed output.
• Implement a small deterministic scorer (RankSixViewProbes, ScoreComponent) that is JudgmentEngine-compatible in shape (explicit scores, reasons, evidence terms, uncertainty, tie-breaking) and prioritizes interior no-hit components, curved analytic families (cylinder/cone/torus/sphere), fallback/tessellated regions, small isolated height bands, centrality, and analytic provenance.
• Bridge map→section by emitting recommended follow-ups for each ranked component: a centroid pointProbe command, two analyze section commands through the component centroid (mapped to supported --xy|--xz|--yz syntax), and a localMap recommendation (marked unsupported until --bounds is added); evidenceBundle contains coarse-map metadata, ranked questions, flattened suggested actions, executedEvidence (empty for A6), limits, and human-readable notes.

Testing

• Ran dotnet restore Aetheris.slnx and dotnet build Aetheris.slnx -f net10.0 --no-restore /m:1 successfully.
• Ran the focused CLI test set with dotnet test Aetheris.CLI.Tests/Aetheris.CLI.Tests.csproj -f net10.0 --filter "AnalyzeMap|Ranked|Probes|Section|EvidenceBundle|Components" and it passed (9 tests).
• Smoke/validation runs: executed aetheris analyze map ... --views six --resolution 8x8 --llm --rank-probes --json on a ruled fixture and confirmed ranked probes/evidence bundle were emitted; exported primitives used in tests included a cylinder and a box.
• Attempted full CTC-01 16x16 six-view ranked-probe run under a 60s cap which timed out (exit 124), validating the documented locality/performance constraint for large dogfood runs (A6 does not auto-execute follow-ups).

---

Codex Task