Normalize coverage path keys for entry-vs-import files

Problem

Coverage reports key files inconsistently: the entry file is keyed by the CLI-spelled (often relative) path, while imported modules are keyed by absolute path (Template.DebugInfo.SourceFile). A file that is both an entry point and an import in the same run receives two separate report records whose hits are never merged — verified during stack #1083's coverage work (layer #1081), where a shared module produced './shared.js' with 4 statement hits plus its absolute-path twin with 5.

Why it was deferred

Normalizing changes the emitted report paths for every consumer (LCOV/Codecov and the JSON format), so it needs a deliberate decision on the canonical form (absolute vs repo-relative) rather than a drive-by fix. See the deferral note in PR #1081.

Sequencing

This now blocks #1041 (coverage scope patterns + thresholds): include/exclude patterns match against exactly these path keys, so implementing scope before normalization would bake the inconsistency into user-facing configuration. Recommend deciding and landing this first.

Where

• Entry-file registration: source/app/Goccia.CLI.SourcePipelineResult.pas (registers the CLI spelling)
• Import registration: bytecode template debug info (absolute)
• Reporters: source/units/Goccia.Coverage.Report.pas (LCOV + JSON emitters would carry the chosen canonical form)

Acceptance

• One record per physical file regardless of entry/import role, with hits merged
• A documented canonical path form, applied consistently in LCOV and JSON output
• A regression test covering the entry+import dual-role case

Read source/app/Goccia.CLI.SourcePipelineResult.pas and source/units/Goccia.Coverage.Report.pas, then trace how bytecode template debug info registers imported files. Decide and document the canonical path form before implementing it consistently in LCOV and JSON. Add a regression test for a file used as both an entry point and an import, verifying one record with merged hits.