refactor(impact): split five modes into impact/ submodules + fixture-based unit tests

[coseto6125]

What

Splits crates/ecp-cli/src/commands/impact.rs (2143 lines — the hottest file in the repo, 23 changes across 300 commits) into commands/impact/{mod,bfs,symbol,baseline,literal,coverage}.rs, mirroring the existing commands/{admin,dev,group}/ submodule convention. Adds the first non-subprocess unit tests for the mode logic.

Why

The file flat-packed five independent "impact" sub-modes (symbol lookup, git-baseline diff, path-literal search, literal-coherence split-brain detection, test-coverage gap analysis) sharing one BFS kernel. High edit frequency + high line count made every change here carry unnecessary review surface and merge-conflict risk.

C1 — pure code motion

• mod.rs: run() dispatch, ImpactArgs, Direction, shared helpers (resolve_min_conf, direction_str, attach_hidden_edges, attach_heuristic_fields, tag_heuristic, emit_hidden_edges_footer, parse_csv_lower), and pub use re-exports so every external call site (main.rs, group/impact.rs, peers_plan.rs, review/aggregate.rs) keeps compiling unchanged.
• bfs.rs: the shared run_bfs kernel plus MergedNodeMeta/MergedEdgeRef merged-space helpers — both symbol.rs and baseline.rs depend on it.
• symbol.rs: impact_by_name, LocalImpact, run_for_symbol, collect_blind_spots.
• baseline.rs: impact_with_baseline, hash_node_lines, head_blob_at.
• literal.rs: build_literal_payload, build_literal_coherence_payload, and the literal-similarity helpers, plus the existing literal_coherence_tests module.
• coverage.rs: classify_symbol, coverage_analyses, build_coverage_json.

Verified pure motion by sorting and diffing the original file's stripped body against the concatenated new modules — every line delta is accounted for by visibility keyword additions (pub(super)), module-qualified call paths, and rustfmt re-wrapping; zero logic drift.

C2 — fixture-based unit tests

Every existing impact_*.rs test spawns the built ecp binary + git init + ecp admin index subprocess. build_payload and run_for_symbol are pub fns taking &Engine, so they're directly unit-testable: serialize a synthetic ZeroCopyGraph to a tempdir graph.bin and Engine::load it — no subprocess, no CLI binary spawn, no .sample_repo or ~/.ecp touch.

New file crates/ecp-cli/tests/impact_fixture_unit.rs covers:

• symbol-mode payload shape on a caller chain (depth/filePath/name fields present)
• ambiguous-name error when two same-named symbols exist with no --file/--kind disambiguator
• --file disambiguator resolving the same ambiguous pair
• direction up vs down producing different, correct result sets
• depth limit stopping BFS before the full chain is traversed
• run_for_symbol wrapping build_payload correctly (direct_symbol_uids/direct_count consistency)

All 6 new tests run in ~0ms total, vs the 200-700ms+ per-test subprocess cost of the existing suite.

Verification

• cargo test -p egent-code-plexus --tests — full suite green (all pre-existing impact_*.rs files pass unchanged, plus the 6 new fixture tests)
• cargo clippy -p egent-code-plexus --tests — clean, no new warnings
• rustfmt --edition 2021 on every touched file — no-op (already formatted)

[github-actions]

ecp impact cache (0 symbols) — internal, used by ecp dev pr-analyze

[]