fix: honor native task names in replacement evidence

[evalstate]

Summary

• Remove the invented result-name/task-ID-name equality in replacement evidence.
• Match native lock names with trial.task_id.get_name(); preserve TrialResult.task_name unchanged.
• Retain native ID/source/configuration, full-source lock coverage, digest/checksum, UUID and multiplicity validation. No prefix stripping, source-specific rules, artifact rewrites, schema/API/UI changes or new resources.
• Preserve prior authorization/completion history additively and record the separately approved narrow matched release. No replacement launch is included.

Harbor ownership review

Checked Harbor dcd0a7ac74b7bd417780d9cb27cd819c7ec82e4e:

• src/harbor/models/task/task.py: public Task.name uses explicit task metadata, otherwise the directory name.
• src/harbor/models/task/id.py: public GitTaskId, LocalTaskId, PackageTaskId.get_name() define task-ID names independently.
• src/harbor/trial/trial.py: result initialization writes task_name=self.task.name and the separate native task ID.
• src/harbor/models/job/lock.py: native lock construction uses task_config.get_task_id().get_name().
• Relevant history: 1d97c20a1 (#1285), metadata-name override; 7f6ae226b (#1552), task-ID-based lock name.

Harbor already owns this distinction: the fix removes a local assumption, not an upstream gap or new wrapper. Existing JobPlan planning/aggregation stays authoritative. No persisted or projected field, configuration alias, parser, scheduler or result format is introduced.

Validation

• Regression demonstrated before the fix: metadata-name aggregation failed while fallback-name aggregation passed.
• Native-built fixtures exercise metadata overrides and fallback names throughout review, parent preflight, cross-language fingerprint validation and aggregation. Additional Git/local/package identity tests and unselected-trial ID/source/lock corruption rejections preserve fail-closed coverage.
• Root Ruff/format/ty: passed; 102 tests, 89.10% coverage; dependency audit passed.
• Agent Ruff/format/ty: passed; 481 tests. Replacement evidence/preflight/aggregation coverage: 98.02%, each module above 85%.
• Node 22.22.0 format/lint/typecheck/build/generated checks and audit: passed; 1,523 unit tests; 75 browser tests.
• Browser runner initially stalled during the unbound-port startup probe; final full suite passed against an explicitly started server from this worktree on an isolated port. No tests changed to bypass this.
• Both Dockerfiles built successfully for linux/amd64.
• Public privacy, diff, normal Slophammer and DRY checks passed.
• Disclosures: required slophammer-py check . --baseline fails because the existing baseline file is absent. Existing agent whole-package coverage gap remains; no threshold was weakened. Native checksum deprecation and existing lint warnings remain.

Full diff and metadata reviewed locally in a separate pass. No independent second reviewer is claimed. No inference, benchmark controls, replacement submission or paid benchmark canary was performed.