diff --git a/apps/elf-eval/src/bin/external_memory_pattern_radar/render.rs b/apps/elf-eval/src/bin/external_memory_pattern_radar/render.rs
index 2823962c..141c576f 100644
--- a/apps/elf-eval/src/bin/external_memory_pattern_radar/render.rs
+++ b/apps/elf-eval/src/bin/external_memory_pattern_radar/render.rs
@@ -23,11 +23,11 @@ pub(super) fn render_summary(cursor: &RadarCursor) -> Result<String> {
 	out.push_str("source_refs: []\n");
 	out.push_str("code_refs:\n");
 	out.push_str("  - apps/elf-eval/fixtures/external_memory_pattern_radar/cursor.json\n");
-	out.push_str("  - apps/elf-eval/src/bin/external_memory_pattern_radar.rs\n");
+	out.push_str("  - apps/elf-eval/src/bin/external_memory_pattern_radar/main.rs\n");
 	out.push_str("related: []\n");
 	out.push_str("drift_watch:\n");
 	out.push_str("  - apps/elf-eval/fixtures/external_memory_pattern_radar/cursor.json\n");
-	out.push_str("  - apps/elf-eval/src/bin/external_memory_pattern_radar.rs\n");
+	out.push_str("  - apps/elf-eval/src/bin/external_memory_pattern_radar/main.rs\n");
 	out.push_str("---\n\n");
 	out.push_str("# External Memory Pattern Radar Summary\n\n");
 	out.push_str("Goal: Preserve the latest weekly ELF external memory pattern radar outcome.\n");
diff --git a/docs/evidence/external_memory_pattern_radar_latest.md b/docs/evidence/external_memory_pattern_radar_latest.md
index cad1348c..4b6c64a8 100644
--- a/docs/evidence/external_memory_pattern_radar_latest.md
+++ b/docs/evidence/external_memory_pattern_radar_latest.md
@@ -14,7 +14,7 @@ tags:
 source_refs: []
 code_refs:
   - apps/elf-eval/fixtures/external_memory_pattern_radar/cursor.json
-  - apps/elf-eval/src/bin/external_memory_pattern_radar.rs
+  - apps/elf-eval/src/bin/external_memory_pattern_radar/main.rs
 related: []
 drift_watch:
   - docs/evidence/external_memory_pattern_radar_latest.md
diff --git a/docs/log.md b/docs/log.md
index 285d64ef..8c0886c4 100644
--- a/docs/log.md
+++ b/docs/log.md
@@ -6,6 +6,14 @@ Read this when: You need to understand why documentation structure changed.
 Not this document: Detailed subsystem history, raw research state, or plan execution
 logs.
 
+## 2026-09-21
+
+- Recorded legacy workspace disposition, historical benchmark claim corrections,
+  and the bounded repair queue. Fixed the radar generator's obsolete entrypoint
+  reference and its checked-in report. Removed the misplaced application-level
+  `--locked` argument from radar tasks; Cargo remains locked. Product behavior
+  is unchanged.
+
 ## 2026-09-20
 
 - Added an exploratory ELF value review using September agent-platform sources and
diff --git a/docs/research/2026-09-21-legacy-workspace-disposition.md b/docs/research/2026-09-21-legacy-workspace-disposition.md
new file mode 100644
index 00000000..efa7f32f
--- /dev/null
+++ b/docs/research/2026-09-21-legacy-workspace-disposition.md
@@ -0,0 +1,100 @@
+---
+type: Research Contract
+title: "Legacy Workspace Disposition"
+description: "Retire obsolete workspaces while retaining evidence and bounded repair candidates."
+resource: docs/research/2026-09-21-legacy-workspace-disposition.md
+status: proposed
+authority: exploratory
+owner: maintenance
+last_verified: 2026-09-21
+code_refs:
+  - apps/elf-mcp/src/app/server/support.rs
+  - packages/elf-service/src/context_pack/assembly.rs
+  - scripts/benchmark_contract/evaluation.py
+  - apps/elf-eval/src/bin/external_memory_pattern_radar/render.rs
+related:
+  - docs/research/2026-09-20-elf-value-review.md
+---
+# Legacy Workspace Disposition
+
+Purpose: Preserve the useful results of retired July and August work without
+restoring obsolete product contracts or benchmark implementations.
+Read this when: Recovering an old workspace or choosing a follow-up repair.
+Not this document: Proof of product safety, a current competitive ranking, or
+approval to change public APIs.
+
+## Disposition
+
+| Former branch | Disposition | Retained value |
+| --- | --- | --- |
+| `xv/benchmark-competitors` | Delivered in PR #381; retire checkout. | Original measured run artifacts and logs. |
+| `xv/benchmark-thin-runner` | Retire transitional runner. | Historical disposition inventory and committed source. |
+| `xv/product-salvage` | Retain commits `2f12e0ff` and `4840feac` as repair candidates. | Response bounds, source checks, lifecycle propagation, and process cleanup. Do not cherry-pick the combined API changes. |
+| `xv/benchmark-contract-refresh` | Retire the experimental implementation. | Historical corrections, source snapshots, and selected test scenarios. |
+| `xy/sag-competitor-benchmark-refresh` | Retire the old competitor runner. | Historical run artifacts and evidence of evaluation defects. |
+| `xy/final-memory-knowledge-plan` | Supersede the July execution proposal in PR #353. | Source-backed memory principles and valid documentation repairs. |
+
+Archives must include ignored artifacts, untracked files, original tracked files,
+and Git history. The local archive manifest records original HEADs, branch names,
+file hashes, and any stash commit used to preserve index and working-tree state.
+Git bundles are verified before checkout removal. Archives remain local: they can
+contain private experiment data and must not be uploaded with a pull request.
+
+## Changes that must not return as cleanup
+
+The contract-refresh experiment changes document read/search/excerpt routes to
+404 responses and removes public search-session and recall-debug routes. It also
+restricts model-facing context to notes and documents. Those changes alter public
+contracts; archive them instead of silently adopting them.
+
+Its source resolver requires the assertion to equal the source quote and accepts
+only a specific v2 source-binding shape. This can reject summaries and existing
+source formats. Reuse source revision and scope checks only after a compatible
+design and migration analysis.
+
+The SAG mem0 adapter builds output claims from `expected_answer.must_include` and
+its expected evidence links. This does not establish independent answer quality.
+Its report status fallback can label out-of-scope-only results as pass while
+listing that state separately. Do not restore this scoring path. The current
+benchmark evaluator separately checks execution, coverage, and score eligibility.
+
+## Historical claim corrections
+
+Retain the July 13 correction to the July 8 benchmark presentation. The old recall
+denominator mixed comparable and non-comparable work. The correction itself says
+that exact original report bytes were unavailable; its numeric examples are
+historical assertions, not independently recomputed measurements. Do not repeat
+those values as a current comparison or use post-hoc intersections as a ranking.
+
+Retain the July 19 correction withdrawing current comparative authority from the
+June 20 and June 27 reports. Fixture-derived results and dated closeout claims do
+not establish present superiority. Preserve original dated reports; link this
+disposition when interpreting them. References in those old corrections to a v2
+experimental contract do not make that contract authoritative today.
+
+## Bounded repair queue
+
+| Priority | Owner | Candidate and acceptance evidence |
+| --- | --- | --- |
+| P1 | MCP transport | Bound response accumulation. Cover declared and chunked oversized bodies. Select a supported limit; the old 1 MiB constant is a proposal. |
+| P1 | MCP transport | Prevent raw upstream error bodies from entering model context. Preserve useful typed error codes and correlation IDs; test text, HTML, malformed JSON, nested JSON, and chunk failures. |
+| P1 | Context Pack | Reject empty or schema-only source identities without requiring every assertion to be a verbatim quote. Preserve supported reference variants with focused tests. |
+| P1 | Service/API | Trace lifecycle state through search, document readback, and Context Pack. Check correction, deletion, scope, and restart behavior against the current contract before porting old tests. |
+| P2 | Local harness | Review child-process reaping from `4840feac`; prove cleanup without changing production lifecycle behavior. |
+
+These are retained candidates, not completed fixes or confirmed exploitable
+vulnerabilities. Old test files depend on an incompatible model-delivery contract;
+port their independent scenarios, not their output schema or route removals.
+
+## Roadmap replacement
+
+The July R0-R9 plan preserves useful provenance, correction, privacy, and cost
+principles. Its broad model-ladder, graph, knowledge-workspace, and dreaming
+expansion sequence is not the current execution queue. Use the September value
+review to design a task-benefit pilot and one cross-agent correction workflow.
+That review remains exploratory and does not itself authorize a new product API.
+
+The old PR also contains valid path and metadata repairs. The radar renderer and
+its checked-in report now reference the existing directory entrypoint. Other
+historical documentation changes require current path checks; old verdicts must
+not be presented as fresh validation.
diff --git a/docs/research/index.md b/docs/research/index.md
index 803a0dac..4a8cbbde 100644
--- a/docs/research/index.md
+++ b/docs/research/index.md
@@ -22,3 +22,4 @@ For legacy research JSON disposition, read
 `docs/evidence/2026-06-18-research-artifact-disposition.md`.
 
 - [ELF Value Review: September 2026](2026-09-20-elf-value-review.md): current native-memory pressure, differentiated value hypothesis, and proposed task-level pilot.
+- [Legacy Workspace Disposition](2026-09-21-legacy-workspace-disposition.md): retired experiments, historical claim boundaries, and bounded repair candidates.
diff --git a/makefiles/research.toml b/makefiles/research.toml
index dc7bc55c..a3dbe00d 100644
--- a/makefiles/research.toml
+++ b/makefiles/research.toml
@@ -10,7 +10,6 @@ args = [
 	"external_memory_pattern_radar",
 	"--",
 	"run",
-	"--locked",
 	"--cursor",
 	"apps/elf-eval/fixtures/external_memory_pattern_radar/cursor.json",
 	"--summary",
@@ -34,7 +33,6 @@ args = [
 	"external_memory_pattern_radar",
 	"--",
 	"run",
-	"--locked",
 	"--cursor",
 	"apps/elf-eval/fixtures/external_memory_pattern_radar/cursor.json",
 	"--out-cursor",
@@ -75,7 +73,6 @@ args = [
 	"external_memory_pattern_radar",
 	"--",
 	"run",
-	"--locked",
 	"--mode",
 	"offline",
 	"--cursor",