feat(compat): add finite AI CLI compatibility harness

[decode2]

Linked issue

Closes #30

• The linked issue has the status:approved label.

Pull request type

Select exactly one option and apply the matching type:* label.

• Bug fix (type:bug)
• New feature (type:feature)
• Documentation (type:docs)
• Refactor (type:refactor)
• Maintenance or tooling (type:chore)
• Breaking change (type:breaking-change)

Summary

• Add the finite, evidence-based compatibility harness for the seven fixed AI CLI and fallback descriptors.
• Report deterministic command discovery and semantic probe evidence in human and JSON forms without promoting unsupported compatibility tiers.
• Wire the focused splice-compat package suite into CI as PR1 of the 15-PR stacked-to-main delivery chain.

Changes

Area or file	Change
crates/splice-compat/	Add fixed descriptors, platform-aware command discovery, semantic probes, report rendering, and native/Windows matrix tests.
Cargo.toml, Cargo.lock	Register and lock the compatibility harness workspace crate.
.github/workflows/ci.yml	Run the focused compatibility package tests in CI.

Test plan

• Ran the focused automated checks: cargo test -p splice-compat (3 passed), cargo fmt --all -- --check (passed), cargo clippy -p splice-compat --all-targets -- -D warnings (passed), and locked metadata (passed).
• Exercised the affected terminal, AI CLI, or desktop workflow: deterministic fixture discovery, startup/input/interrupt/liveness evidence, fallback behavior, and human/JSON rendering all passed.
• Verified relevant Windows, Linux, or WSL behavior, or documented why it is not applicable: native Linux tests passed; the complete Windows GNU matrix ran under Wine with 3 passed, including PATHEXT and direct .CMD discovery. WSL-specific integration is intentionally deferred beyond PR1.

Differential baseline and warnings

• cargo test --workspace exits 101 on both this candidate and exact base eab3555003d84dd2ddc721eaf661fe11a38e0842, with the same missing desktop icon and unavailable Arc<PtySession>::close failures. These failures are pre-existing in untouched desktop/PTY paths and are not introduced or worsened by PR1.
• Review findings R3-001 and R3-002 remain informational: caller-supplied evidence is reused per discovered CLI, and false semantic probes render as skipped. Neither behavior promotes tiers or invalidates the PR1 scenarios.
• Native review lineage review-2b2b0b65ee634889 approved candidate tree fbeb78d43bebd5a589b4b0fe7457f676bd2bb2fc with no blockers or critical findings.

Chained pull request context

• Strategy: stacked PRs to master
• Chain position: PR1 of 15
• Dependency: master -> PR1 [current] -> PR2 (planned) -> ... -> PR15 (planned)
• Previous PR: Not applicable; this is the independently mergeable foundation slice.
• Next PR: PR2, planned scope only; it will target master independently after PR1 delivery.
• Review this PR first: the finite descriptor matrix, platform command discovery, semantic evidence model, renderers, and focused CI boundary.
• Intentionally out of scope: PTY implementation changes, platform adapters, desktop/UI integration, packaging, broader documentation/marketing claims, WSL-specific integration, and all PR2-PR15 work.
• Review budget: 398 additions + 1 deletion = 399 changed lines, within the 400-line limit.
• Rollback: revert this one commit to remove crates/splice-compat/, its workspace/lock registration, and the focused CI command without removing unrelated behavior.

Contributor checklist

• I searched for duplicate issues and pull requests before starting.
• I linked an approved issue using Closes, Fixes, or Resolves.
• I selected exactly one pull request type and applied exactly one matching type:* label.
• I completed every relevant template field and included reproducible test evidence.
• I updated documentation when behavior or compatibility changed, or documented why broader documentation is deferred beyond this harness-only slice.
• My commits follow Conventional Commits.
• My commits contain no AI attribution, including Co-Authored-By trailers.