fix: analyze include roots as fragments

[evanbrobertson]

Closes #158.

What changed

• classify an explicitly opened .i root as an include fragment across the shared pipeline, CLI, and LSP
• reclassify only includer-dependent unresolved names as external while preserving qualified class and literal RUN workspace misses
• withhold LINT0002, LINT0005, and LINT0006 when their whole-compilation-unit counts are unavailable
• keep parse, preprocessing, schema-backed, type, and other locally provable diagnostics active
• expose the reduced coverage through the CLI and analyze envelope
• add shared parity coverage for root-file identity, including explicit browser capability handling

Why

An .i file is a textual fragment. When opened directly, declarations, imports, scopes, reads, and writes supplied by its includer are absent. Treating that fragment as a complete compilation unit produced large waves of false undefined-symbol, unused-variable, and dead-store findings in the editor.

The root cause was not a different lint implementation in the LSP: every client shares the same pipeline, but the editor can feed that pipeline an .i as a root while directory discovery deliberately feeds the CLI only complete root extensions.

Impact

Directly opened include fragments now provide useful but explicitly reduced diagnostics. The same .i expanded through a .p, .w, .cls, or .v remains part of an ordinary complete-unit analysis with all rules enabled.

On the reported include, the merge-base produced 57 false lint findings (44 undefined symbols, 5 unused variables, and 8 dead stores); this branch produces none. A deterministic sample of 50 include roots fell from 684 findings to one locally provable type mismatch. An A/B over 10,482 ordinary roots produced byte-identical JSON and stderr, confirming that complete-unit behavior did not move.

Verification

• cargo fmt --all -- --check
• cargo check --workspace --all-targets
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace
• deterministic merge-base versus candidate corpus A/B