Recover from damaged build artifacts instead of failing every build

[mkarlesky]

A single damaged file under build/ could leave every later build failing until a clobber. Fault-injection stress testing found three such cases. This PR makes each one cost a rebuild of the affected step instead.

Failures

• A truncated or emptied cached test result (build/test/results/<test>.pass) was reused as-is. The summary report then raised undefined method `size' for nil or undefined method `empty?' for nil.
• A truncated .d file can end in a directory path. The directory was tracked as a dependency, and hashing it raised Is a directory @ io_fread before the recompile that would have rewritten the file.
• A deleted build-directives cache (build/test/preprocess/build_directives/<test>/) raised Could not find YAML file …_source_files.yml. An emptied one was worse: it loaded as no directives at all, silently dropping TEST_SOURCE_FILE() sources. That surfaced only as undefined reference at the next relink.

Cause

Two of the three share a cause: a freshness record vouched for an artifact it never checked. Test execution was skipped when <test>.fixture_run was fresh, then reused the result file. Directive parsing was skipped when the preprocessed test file was fresh, then loaded the cache. Neither artifact was tracked, so damage to it was trusted indefinitely.

The .d case was the tracker counting any existing path as a dependency, directories included.

Fix

The dependency tracker records outputs. register accepts outputs:: files a step produces and a later run trusts. mark_fresh records their hashes in the target's cache entry. stale? reads them back from that entry, not this run's registrations, so a step registers outputs only after producing them. A missing or altered output makes the target stale, and the step reruns. An entry without outputs is unchanged, so existing caches need no schema bump.

Only a regular file counts as present for a target, a dependency or an output. A directory named by a damaged .d file reads as missing, so the object recompiles and gcc writes a fresh .d.

Call sites:

• Test execution records the result file it wrote.
• Test file preprocessing records its build-directives cache. It now marks its target fresh only after storing the cache, so a failure partway leaves the target stale.
• A review of every mark_fresh site found the same gap elsewhere. Directives-only preprocessing now records its compacted file, and Partial preprocessing its includes list and full expansion, all of which a skipped run reuses.
• Mock and Partial implementation headers were tracked as dependencies of the file written beside them, with comments noting they weren't really antecedents. They are now outputs. Detection is the same, and the workaround comments are gone.

Changes

• lib/ceedling/dependencies/dependency_tracker.rb: register( outputs: ), the regular-file rule, and output checks in stale?, mark_fresh and diagnose, sharing one presence-and-hash check with dependencies.
• lib/ceedling/dependencies/dependinator.rb, dependency_cache_store.rb, README.md: the passthrough, validation of outputs when present, and documentation.
• lib/ceedling/test_invoker/test_build_executor.rb, test_build_setup.rb, partials_manager.rb: the call sites above.
• lib/ceedling/generators/generator.rb: its comment on cached results reflects that damage is now caught first by staleness.

Tests

• Unit: directory dependencies, output recording and checking, compatibility with entries that have no outputs, diagnose, cache validation, the passthrough, and each call site's registration order.
• Integration: on real files across two tracker runs, a truncated, emptied or deleted output reads stale, and a .d file naming a directory neither raises nor reads fresh.
• System, new spec/system/build_artifact_recovery_spec.rb: damages one artifact after a passing build and expects the next build to pass. It covers a truncated .pass, an emptied .pass, a .d naming a directory, a deleted directives cache, an emptied directives cache followed by a code change, and a truncated mock header. Against next_version, five of the six fail with the original errors. The mock-header case was already handled and guards the consolidation.

An untouched second build still skips every stage. Units pass 4018/0 and integration 304/0, along with the delta-build, gcov, Partials and include-handling system specs. RuboCop is clean. The stress rig reran both original seeds with 60 random cycles each and found no build that failed to recover, across 504 builds.

Not covered

A truncated .d file that only loses entries, rather than naming a directory, under-tracks headers without failing. That is a pre-existing limitation of dependency files and is not sticky.