Compact debug source paths

Issue #704 implements the first size opportunity from the optimization survey. Repeated source paths now live in a per-module table. DEBUG entries keep all function, block, instruction, UTF-16 range and inline-parent fields; source line/column records are unchanged. Rust shares path ranges in admitted bytes; Kotlin shares pooled text objects. Small tables keep inline paths when the extra table and directory entry would increase size. The writer compares aligned physical costs before constructing legacy rows, so compact output can fit a limit that legacy metadata exceeds.

Baseline parent: 6528a680; baseline VM: 18103f01963874510fa25d4396a2dc3e7ee47640. Updated VM: 190479a. The compiler freshly regenerated all six builtins and the two larger collection artifacts on the same workstation. Before byte sizes and SHA-256 values are preserved in the earlier survey. KiB = 1024 bytes.

The editor falls from 220116 to 181788 bytes; the large mapNotNull example falls from 561176 to 397752 bytes. These are physical executable sizes, without compression. No interpreter speedup is claimed by this change.

Follow-up: compact GC root ranges and combined size results. The table above remains the measured path-pooling stage.

Equivalence and compatibility

Expanding the new path pool into legacy DEBUG records, removing DEBUG_PATHS and rebuilding the container produces exactly the earlier artifact SHA-256 for all eight files. This checks every semantic section, every diagnostic entry, source position, inline parent and all other payload bytes against the recorded baseline. It does not merely compare counts or source filenames. The updated native public verifier accepts all eight compact files.

The critical, nonsemantic DEBUG_PATHS section (0x0111, flags 1) selects the compact DEBUG representation. Readers without support reject it; updated readers also accept old inline-path artifacts. Container format 3.0, semantic Runtime ABI 1.15 and native C ABI 20 remain unchanged. The frozen baseline Runtime 0.20.0 benchmark binary rejects compact boot with BadSection: unknown required or semantic section, as intended. See the active ABI reference for encoding details.

Published Runtime 0.20.0 bundles do not contain this reader. Use source-built natives from the updated VM. A future immutable Runtime release and parent bundle-pin update are needed before packaging compact compiler output with published natives. No published pins were advanced to nonexistent assets, and this work does not claim release readiness.

Verification

• cargo test --manifest-path host/compukter-vm/Cargo.toml --locked --offline: 545 passed, 10 ignored, across unit, integration and documentation tests. Legacy golden fixtures remain unchanged. New cases cover malformed/duplicate/empty pools, first-use order, unused entries, orphan pools, indices, UTF-8, canonical paths, widths, parents, flags, scopes and combined debug-byte limits. The computer process OOM test exercises a compact child artifact and checks at entry (src/main.kt:3:8) while retaining its legacy-offset and no-debug cases.
• cargo fmt --manifest-path host/compukter-vm/Cargo.toml -- --check and cargo clippy --manifest-path host/compukter-vm/Cargo.toml --locked --offline --all-targets -- -D warnings.
• :compiler-artifact:check: 161 JVM tests passed; reader/writer round trips, deterministic bytes, shared multibyte paths, inline parents, semantic identity, legacy fallback, malformed pools and bounded compact output.
• All six builtin generator determinism checks plus testKotlinMapNotNullVmConformance, testKotlinMutableListVmConformance and testKotlinExceptionsVmConformance: compiler output executes in the updated Rust VM.
• Independent public verify_artifact admission of the six builtins and two large collection files.
• Read-only equivalence verification and section accounting across all eight files. No full-checkout, GameTest, cross-platform native-bundle or release gate is claimed.

Exact sizes and equivalence hashes, section inventories, verification commands.

Reproduce after regenerating the inputs, from the repository root:

python3 docs/benchmarks/debug-paths-704-2026-10-07/verify_legacy_equivalence.py
python3 docs/_tools/analyze_artifact_size.py modules/common/compiler-k2/build/generated/system/*.cpkt

The equivalence script reads artifacts and the archived baseline, expands only in memory, and checks the exact baseline hashes before printing JSON. It does not write executable artifacts or substitute for native verification.