Java-to-Rust Migration Testing

Prove observable compatibility and Rust safety properties, not test volume. Treat the complete source test suite as a non-negotiable compatibility floor: migrate 100% of its test cases without weakening inputs or assertions, copy every test fixture/data/script byte-for-byte into the target repository, and make both implementations produce the same per-case observable result. Build three explicit test ledgers in this order:

Mandatory verification structure

• Require exact Java/Rust path parity; every MISPLACED or missing file-map.csv target blocks completion.
• Treat E0583/E0761/E0405/E0425/E0659 and broken fixture paths as structural failures after file moves.
• Treat up to 500 physical lines as normal, review authored .rs files from 501–800 lines for cohesion, and block files above 800 lines; split suites by contract or behavior family rather than arbitrary ranges.
• Allow focused unit tests in #[cfg(test)] modules beside their implementation. Put public-boundary, cross-module, differential, host, load, and whole-project tests in tests/ or a non-published test/testkit crate.
• Run path audit, fmt, default/all-feature check, Clippy, complete tests, golden/live differential, migration audit, and the project coverage gate without using any one result as parity proof.

Read Directory parity verification and Migration verification SOP.

1. SOURCE_PARITY — map every source test/case and every source test asset; no missing, blocked, or not-applicable row permits a completion claim.
2. RUST_OBLIGATION — test risks introduced by the Rust implementation and replacement components.
3. VALUE_ADD — add tests justified by uncovered behavior, plausible defects, incidents, mutation survivors, or hostile inputs.

SOURCE_PARITY has two mandatory sections: TEST_CASE and TEST_ASSET. TEST_CASE preserves the source inputs, assertions, exception/error category, ordering, state, side effects, and cleanup. TEST_ASSET preserves the source relative path, target path, and SHA-256 of every copied fixture, script, corpus, golden file, and data file. Coverage should rise because meaningful contracts are exercised. A percentage is not the design input and 100% is not migration proof.

Scope and routing

Use this skill for migration-specific test planning, implementation, audit, and acceptance.

This is the Rust testing adapter of the shared target-language migration profile. Keep source-test dispositions, exact-asset rules, full differential outcomes, and completion semantics equivalent to zig-java-migration-testing and future Kotlin/Swift adapters; vary only Rust test/build syntax and Rust risks.

Route:

• module/object/component migration planning to rust-java-migration;
• general Rust unit, integration, doctest, and fixture mechanics to rust-testing;
• async model checking to rust-concurrency;
• profiling and benchmark construction to rust-performance;
• unsafe and FFI validation to rust-unsafe-ffi;
• web threat modeling to rust-web-security.

An audit-only request does not authorize deleting or rewriting tests. Preserve existing behavior and unrelated dirty work. Treat heuristic findings as review candidates.

Required inputs

Resolve or mark unknown:

• pinned Java and Rust SHAs, dirty state, toolchains, profiles, features, targets, and generated-code boundaries;
• the authoritative workspace-topology decision and actual paths of the target production and <project>-test packages; do not assume a crates/ directory;
• the complete source test roots and runner configuration, including fixtures, examples, scripts, test resources, parameter sources, dynamic factories, and test-support code; the source repository defines the denominator and exclusions cannot be introduced merely to make parity reach 100%;
• Java test runner and Rust test runner, including parameterized and dynamic test behavior;
• source coverage scope/tool/report and comparable Rust coverage scope/tool/report;
• available oracle: source only, executable Java tests, golden exporter, packaged artifact, live service, or standards suite;
• the current authoritative object ledger generated from the same Java/Rust baselines; ignore historical-design appendices when reading current states;
• required hosts, real dependencies, concurrency/load model, security boundary, and rollback mechanism;
• deterministic normalization rules for time, identifiers, paths, map order, locale, float precision, and scheduling.

Never use an unpinned remote artifact as the compatibility oracle.

Evidence labels

Use these labels without promotion:

A copied test name, green Rust suite, or 100% line report does not prove V3_GOLDEN_DIFF.

Evidence levels never replace object states. A green V1_RUST_LOCAL suite does not turn MISSING, MISPLACED, STUB, PARTIAL, or UNVERIFIED into IMPLEMENTED. DEPENDENCY_REUSED additionally requires an exact pinned upstream symbol and local integration test; PLATFORM_NA requires platform evidence rather than a passing test that skips the behavior.

SOP

1. Freeze the verification baseline

Record exact SHAs, commands, tools, profiles, features, targets, test counts, ignored/flaky tests, coverage exclusions, and existing evidence artifacts. Record Java and Rust coverage separately before comparing them.

Read the current fact region of the authoritative object table before designing tests. Stop at historical-design-appendix-start; old appendix statuses are context only. Record counts for all strict object states. If any incomplete state remains, the acceptance report must say “module migration incomplete” regardless of test results.

Use CodeGraph when indexed to trace each source test through its production entry, collaborators, side effects, and Rust counterpart. Text similarity is insufficient for overloaded methods, registries, interceptors, dynamic dispatch, cleanup, and async paths.

2. Inventory source and target tests

Create one row per Java test method and one row per concrete parameterized, repeated, template, or dynamic case. Include disabled tests; a disabled source test remains a required migrated test and its disabled reason is preserved separately. Create one TEST_ASSET row per file in every source test resource, fixture, script, corpus, and data root. Copy source assets without editing them; place target-generated or normalized derivatives beside the immutable copy.

Use dispositions:

MIRRORED, ADAPTED, SPLIT, and MERGED_APPROVED are complete only when every source case remains identifiable, all preservation flags are true, every target test exists, and golden/live comparison records MATCH. ADAPTED permits Rust-native harness mechanics, not a weaker contract. SPLIT and MERGED_APPROVED may reorganize tests but may not remove a case or assertion. Do not map by test name alone.

Resolve SKILL_DIR to this skill directory and run from the migration repository:

python3 "$SKILL_DIR/scripts/audit_migration_tests.py" \
  --java-root ../java-project/source-module \
  --rust-root <selected-target-crate-or-workspace> \
  --object-ledger docs/source-module/对象级对照表.md \
  --parity-manifest docs/source-module/source-test-parity.json \
  --java-test-assets-root src/test/resources \
  --fail-on-incomplete

The report inventories tests, flags weak signals, verifies every source-test manifest row, checks target test files, hashes exact asset copies, and refuses a completion gate for any object, test, asset, run, or differential gap. Additional non-standard asset roots must be passed with repeated --java-test-assets-root. The script validates recorded preservation evidence; it cannot infer semantic mappings or authorize deletion. It warns on authored files above 500 physical lines and blocks those above 800; inline unit tests are valid Rust organization and are not reported as structural violations.

3. Implement the SOURCE_PARITY ledger

For every source row:

1. Trace the protected Java contract and production call path.
2. Port the fixture and assertions, not merely the method name.
3. Preserve valid, boundary, failure, state-transition, and side-effect cases.
4. Run the source and target suites from pinned baselines and retain raw per-case results. A mirror without Java output is useful during implementation but does not satisfy the final source-parity gate.
5. Compare each source case through a pinned golden or live differential oracle; require MATCH, zero harness failures, and zero not-run cases.
6. Record the Rust test, preservation flags, evidence, commands, artifacts, and any divergence.

Missing, blocked, or not-applicable source tests remain visible and block a migration-complete conclusion. Source tests are a compatibility floor, not the complete Rust plan.

Whole-project migration test module

For every repository/product-level migration completion claim, create one non-published workspace package as the whole-project acceptance authority. It is especially important for multiple crates, bindings/adapters, or a source system/templatesuite. Name it <project>-test by default; keep the final directory name and Cargo package name identical, set publish = false, and run it explicitly in CI. Its parent path follows the recorded workspace topology: a small root-flat workspace uses <project>-test/; a hybrid may keep it at root or in a real test family; a contained workspace may use crates/<project>-test/ or crates/tests/<project>-test/. Local tests inside production crates prove their own parser/type/API/binding behavior; they do not replace the whole-project module's source-suite replay, cross-crate workflows, real copied assets, golden/live differential comparison, or aggregate result artifact.

For FreeMarker use this boundary:

freemarker/          # Rust engine local unit/integration tests
freemarker-pyo3/     # Python binding local/packaging tests
freemarker-test/     # complete Java templatesuite + cross-component acceptance

The freemarker-test package must be a workspace member, depend on public surfaces under test, and remain outside crates.io publication. Its complete gate is every source case MATCH, with no threshold pass count, skipped capability, or ignored case. Read Migration verification SOP for naming, ownership, layout, and CI rules.

The test package's semantic role is mandatory; one physical parent directory is not. Never relocate it solely to imitate the Java test-module tree or a generic Rust example.

Do not use one test per object as a substitute for one real file per source object. A test that reaches a re-export, compatibility facade, or merged type does not cure MISPLACED/MISSING. Tests validate semantics only after the layout and object boundary are factually present.

Treat 500 physical lines as the suite-cohesion review threshold and 800 as the authored-file blocker. Split by contract, behavior family, fixture family, adapter, or verification layer while preserving every source case ID and assertion. Keep compact private-behavior unit tests in #[cfg(test)] modules; put reusable integration support in tests/common/ or a dedicated testkit crate. Enable clippy::too_many_lines to review functions over its 100-line default.

4. Implement the RUST_OBLIGATION ledger

Add applicable tests created by the target design:

These tests need not exist in Java because they protect the Rust implementation's correctness.

For a DEPENDENCY_REUSED row, execute the local adapter against the exact declared upstream symbol/version or commit. The dependency's own unit tests, documentation examples, or a similar capability name are not local integration evidence. Assert the source contract's ordering, errors, lifecycle, cancellation, and metadata that cross the adapter.

5. Implement the VALUE_ADD ledger

Add a test only when it has a reason such as:

• missing boundary or branch revealed by coverage;
• surviving meaningful mutant;
• property/invariant over a broad input space;
• malformed or hostile input found by fuzzing;
• production incident or bug regression;
• concurrency interleaving, load, resource leak, or long-soak risk;
• compatibility behavior absent from the Java suite but required by docs/protocol;
• real-host or rollback behavior that mocks cannot prove.

For each test, record the concrete bug it should catch. Prefer a small parameter/branch matrix over “one test per type/token/method” count targets.

6. Use observable assertions

Strong tests observe exact public behavior:

• value, bytes, order, state transition, call count, side effect, resource release, or typed error;
• Display, Debug, serialized error report, logs/tracing fields, transport body, and Error::source() independently where exposed;
• cache hit/miss/eviction metrics or backend-call counts when claiming cache behavior;
• body/trailer/backpressure and disconnect cleanup when claiming streaming lifecycle.

Weak patterns requiring review:

• let _ = result, unused parsed AST, or a test that accepts success and failure; note that during cleanup, converting a genuinely dropped Result to let _ = expr; preserves semantics and is not itself weak evidence — it only becomes weak when let _ = is the sole observation of the behavior under test;
• parse-only checks named as semantic/evaluation tests;
• only is_ok() or is_err() when value/error category matters;
• a cache test that observes only final values;
• tests of fixture Clone, Debug, constants, constructors, or type existence without a public contract;
• duplicate “coverage burst” tests with no distinct risk;
• tests that merely prove todo!() or unimplemented!() panics.

Read Vernal case study and test-value rubric before cleanup.

7. Build differential fixtures

Use a versioned case format such as JSON Lines:

{"schema":1,"case_id":"empty-name","input":{"name":""},"expected":{"kind":"error","code":"INVALID_NAME"}}

Retain Java and Rust raw outputs separately. Pin exporter and implementation SHAs, seeds, environment, and normalizer version. Compare success values, error categories, side effects, order, and wire bytes as applicable. Do not normalize unexpected fields away.

Differential comparison is a full source-suite gate, not a representative sample. The final manifest must record one MATCH per concrete source case and suite-level Java, Rust, and differential runs with PASS, zero mismatches, zero harness failures, and zero not-run cases. A temporary subset may guide development but must be labeled partial and cannot satisfy completion.

8. Reuse conformance suites for adapters

Put shared assertions and failure fixtures in a testkit. Each adapter must provide real native observations. Run the same identity, scope, lifecycle, error, cancellation, streaming, and cleanup contracts for every implementation, then add adapter-native tests.

Prefer event-driven synchronization and bounded timeouts over fixed sleeps. Assert the resource is open while a stream is active and closed after completion, disconnect, timeout, cancellation, or panic.

9. Compare coverage without gaming it

Coverage comparison is valid only when scopes are documented and reasonably comparable:

• production files/modules included;
• generated code and approved exclusions;
• line/region/branch semantics of each tool;
• features, targets, test types, and profile;
• source and target baselines.

Acceptance order:

1. every source test and concrete case has a lossless target implementation;
2. every source test asset has a byte-identical target copy verified by SHA-256;
3. both complete suites pass and every source case has differential MATCH;
4. every high-risk source contract has adequate evidence;
5. every applicable Rust obligation is tested;
6. meaningful mutants and uncovered branches are reviewed;
7. comparable Rust coverage exceeds the Java baseline if the project requires it.

Do not weaken assertions, duplicate tests, exclude difficult files, or add trivial getters to reach a number. A user-mandated 100% gate may be enforced, but report what it proves and what it does not.

10. Audit test value before removing or merging

Score each test manually:

Classify KEEP, IMPROVE, MERGE, or REMOVE_PROPOSED; require review before deletion. Coverage loss alone may justify replacement but not retention of a meaningless assertion.

11. Run layered gates

Run applicable gates from cheapest to most diagnostic:

cargo fmt --all -- --check
cargo check --workspace --all-targets --all-features
cargo test --workspace --all-features
cargo clippy --workspace --all-targets --all-features -- -D warnings
cargo test --doc --workspace

Then run targeted compile-fail, platform/MSRV, differential, real-dependency, real-host, model/Loom, Miri/sanitizer, mutation, property/fuzz, load/soak, security, and rollback gates. Record exact command, environment, result, and artifact.

Run every gate at the workspace MSRV: a default rustc older than the workspace rust-version fails with rustc X is not supported by the following packages rather than testing your code — use rustup run <ver> cargo ... and never silently lower the manifest. Use --keep-going so one failing crate does not hide the remaining errors (errors surface in layers; fix the first layer and re-run). Require both default and --all-features check runs to reach zero warnings: feature-gated crates and code paths only compile under --all-features, and a check gate with warnings must be reported as a gap, not a pass.

Use the project-specific coverage command. For mutation candidates:

"$SKILL_DIR/scripts/run_mutation_test.sh" <selected-target-crate>

Interpret survivors individually; do not impose one universal mutation score.

12. Report acceptance honestly

Report separately:

• source-test lossless implementation and per-case differential coverage;
• source behavior evidence by V0–V4;
• Rust-obligation completion;
• value-add tests and defect/risk rationale;
• line/branch/region coverage with comparable scope;
• mutation, property, fuzz, concurrency, load, security, host, and rollback evidence;
• failures, flaky/ignored tests, stubs, exclusions, missing targets, and external boundaries.
• current object-state counts and the explicit list of all MISSING/MISPLACED/STUB/PARTIAL/UNVERIFIED blockers;
• compiler and Clippy warnings, ignored/doctests, feature/target gaps, and tests that were not executed;

Red lines

• Do not call mirrored tests differential.
• Do not replace source-test disposition with raw test-count parity.
• Do not exclude, disable, weaken, merge away, or mark a source test not applicable to manufacture 100% parity.
• Do not rewrite copied source fixtures or data in place; retain a byte-identical copy and generate target-specific derivatives separately.
• Do not accept two independently green suites as result parity; require per-case Java/Rust golden or live comparison.
• Do not substitute local tests in production or binding crates for the non-published <project>-test whole-project acceptance package, and do not use pass thresholds, skips, or ignored cases as its completion gate.
• Do not hardcode <project>-test under crates/ or at repository root without checking the migration roadmap's project-driven workspace topology.
• Do not write tests solely to increase coverage or file count.
• Do not keep an authored .rs file above 800 physical lines; review every file above 500 and split weakly cohesive suites.
• Do not compress code, remove useful comments, or split arbitrary line ranges to evade a size gate.
• Do not place integration, differential, host, load, or whole-project suites in production modules; focused inline unit tests remain valid.
• Do not call parse success semantic equivalence.
• Do not accept generic is_err() when the error contract is observable.
• Do not auto-delete tests from names, body length, or heuristics.
• Do not hide mismatches through broad normalization or snapshot regeneration.
• Do not replace real dependency/host tests with mocks when the contract crosses that boundary.
• Do not use fixed sleeps as the only async coordination mechanism.
• Do not count production stubs as implemented because their tests compile.
• Do not mark a module complete from Cargo/JUnit test totals, coverage, or a green CI job while the authoritative object ledger has any incomplete state.
• Do not run a completion gate without a source-test parity manifest. Missing test mappings, incomplete parameter expansion, false preservation flags, missing target files, asset hash differences, non-MATCH results, harness failures, or not-run cases all block completion.
• Do not read completion states from a historical appendix or stale duplicate migration document.
• Do not call dependency reuse verified from upstream tests or semantic similarity; require the exact dependency symbol and a local integration test.
• Do not report the check gate as passed while cargo check (default or --all-features) still emits warnings, or when the run used an unsupported toolchain instead of the workspace MSRV.

On-demand resources

• Migration verification SOP
• Directory parity verification and post-migration checks
• Cross-language migration profile
• Test categories
• Test-value rubric
• Vernal positive and negative examples
• Migration test ledger template
• Machine-readable source-test parity manifest
• Worked audit report
• scripts/audit_migration_tests.py — Java/Rust test inventory and weak-signal audit
• scripts/run_mutation_test.sh — mutation-test wrapper

Completion criteria

• Every Java test and every concrete parameterized/dynamic/repeated case in the frozen source test roots has a target-language implementation and source trace; the source denominator is 100% accounted for without completion exclusions.
• Every source fixture, resource, script, corpus, golden file, and test data file has a byte-identical copy in the target repository with matching SHA-256.
• Every mapped case attests preservation of contract, inputs, assertions, fixture state, and cleanup; split/merged/adapted forms lose nothing.
• The complete pinned Java and Rust suites pass, and the complete golden/live differential report contains only MATCH with zero harness failures and zero not-run cases.
• The <project>-test package is a workspace member with publish = false, lives at the path selected by the recorded workspace topology, exercises public component surfaces, and owns the complete suite/differential command plus aggregate artifact; local crate tests remain subsystem evidence.
• The authoritative current object ledger was checked, its baselines match the test run, and no incomplete object state was hidden by the test summary.
• Every high-risk contract has an oracle, evidence label, and result.
• Applicable Rust ownership, async, error, serialization, feature, adapter, and unsafe obligations are tested.
• Added tests name a distinct risk or plausible defect.
• Every authored .rs file is at most 800 physical lines; files above 500 and functions above the Clippy 100-line signal have recorded cohesion reviews.
• Focused unit tests may be colocated in #[cfg(test)] modules; integration, differential, host, load, and whole-project suites live under tests/ or a non-published test package.
• Coverage scopes are comparable and any numeric gate is reported as a signal, not parity proof.
• Stubs, warnings, flaky/skipped tests, missing platforms, real-host gaps, and unverified boundaries remain visible.
• A module completion claim is emitted only when its object ledger, source-test case ledger, source-asset ledger, complete differential result, Rust obligations, and required host/non-functional gates all permit it.