diff --git a/UPSTREAM_CHERRY_PICK_CANDIDATES.md b/UPSTREAM_CHERRY_PICK_CANDIDATES.md
new file mode 100644
index 000000000..e10faa4f0
--- /dev/null
+++ b/UPSTREAM_CHERRY_PICK_CANDIDATES.md
@@ -0,0 +1,290 @@
+# Upstream cherry-pick candidates
+
+A read-only enumeration of high-value commits in `gsd-build/gsd-2` (`upstream/main` at `fec206dda`, 2026-04-28) that are not in `singularity-ng/singularity-foundry/main` (at `b24f426f2`, 2026-04-29).
+
+Total upstream-only commits: 4,589. This list is the **high-leverage subset** worth cherry-picking. Skipping the bulk of small/internal commits.
+
+Clusters are roughly ordered by "take this first." Each cluster lists candidate SHAs with one-line context.
+
+---
+
+## A. `/gsd eval-review` feature (~17 commits)
+
+A new command for milestone-end evaluation review, with frontmatter schema and integration tests. Single coherent feature; cherry-pick as a block.
+
+```
+979487735 feat(gsd): add EVAL-REVIEW frontmatter schema module
+6971f4333 feat(gsd): add /gsd eval-review command handler
+a2f8f0e08 feat(gsd): register /gsd eval-review in catalog and ops dispatcher
+83bcb054c feat(gsd): emit pre-ship soft warning on EVAL-REVIEW status
+a686d22cb test(gsd): add /gsd eval-review integration suite
+087cd6a0f docs(gsd): add /gsd eval-review user spec, drop ADR-011 references
+176fa5c99 fix(gsd): include eval-review in /gsd help full output
+bc8e17cd6 refactor(gsd): strip PR/issue references from eval-review code comments
+35f5e2b57 docs(gsd): label fenced code blocks in eval-review.md (markdownlint MD040)
+d2bf7e7d0 docs(gsd): vary lead phrasing in eval-review Related section
+f2206dac3 fix(gsd): degrade AI-SPEC.md read failure to a marker instead of throwing
+62207fc8a fix(gsd): clamp computeOverallScore to MIN_SCORE..MAX_SCORE
+c0e778b2f fix(gsd): handle UTF-8 multi-byte chars at the truncation boundary
+090c02d31 fix(gsd): three CodeRabbit findings — control flow, marker budget, Windows test
+8931209c5 fix(gsd): bound eval-review reads to cap and surface AI-SPEC errors
+ac71c03b7 fix(gsd): three CodeRabbit findings on eval-review prompt and budgeting
+e111ed88f Merge pull request #5118 from NilsR0711/feat/eval-review-v2
+18ce71551 fix(gsd): allow review-tier subagent dispatch from validate-milestone
+089be6f07 Merge pull request #5099 from jeremymcs/fix/validate-milestone-dispatch-policy
+```
+
+Effort: ~2 hours. Touches: `src/resources/extensions/sf/eval-review*`, command catalog, help text.
+
+---
+
+## B. `agent-session` / `agent-end` transitions (4 commits — critical)
+
+These fix real session-transition bugs. Should take regardless of other choices.
+
+```
+71114fccf fix(agent-session): guard synthetic agent_end transitions
+6d7e4ccb5 fix(agent-session): skip idle wait after agent_end
+e3bd04551 Fix session transition during agent_end
+c162c44bf Fix agent_end session switch handoff
+```
+
+Effort: <1 hour. Likely lands cleanly.
+
+---
+
+## C. claude-code-cli permission persistence (3 commits)
+
+Always-Allow for non-Bash tools didn't persist; fix + tests.
+
+```
+a88baeae9 fix(claude-code-cli): persist Always Allow for non-Bash tools
+1cce8ae38 test(claude-code-cli): cover empty permission suggestions fallback
+bf1d8aad0 Merge pull request #5096 from jeremymcs/fix/always-allow-non-bash-tools
+```
+
+Effort: <1 hour.
+
+---
+
+## D. Worktree TUI commands (2 commits)
+
+Adds `worktree list|merge|clean|remove` to the TUI dispatcher.
+
+```
+2361ceeb1 feat(gsd): add worktree {list,merge,clean,remove} commands to TUI dispatcher
+325aae489 Merge pull request #5055 from jeremymcs/feat/worktree-tui-commands
+```
+
+Effort: <1 hour. Touches: `src/resources/extensions/sf/worktree-command*.ts`.
+
+---
+
+## E. Worktree path safety + normalization (~12 commits)
+
+A series of fixes hardening worktree path handling against injection, self-merge, dirty handling, cwd anchoring. Ship all together.
+
+```
+0fdacd524 Merge pull request #5062 from jeremymcs/fix/worktree-path-injection
+16f025a0e Merge pull request #5051 from jeremymcs/fix/worktree-root-normalization
+84a383f51 Merge pull request #5041 from jeremymcs/fix/5024-prevent-self-merge
+f6d51492f fix(gsd): normalize worktree project roots
+cf9927a1a fix(gsd): normalize auto worktree loop roots
+17fce6461 fix(gsd): harden worktree dirty handling
+ca7a0bc14 fix(gsd): anchor subagent dispatch to canonical worktree path
+de73fb43d fix(gsd): stop dispatch on cwd anchor failures
+4aff417ee fix(gsd): anchor cwd at project root in mergeAndExit (closes #5079)
+fabecd488 fix(gsd): harden worktree dispatch cwd handling
+7cfa24af6 fix(gsd): anchor cwd without cwd guard
+13426f8cb fix(gsd): normalize self-merge ref guard
+82bcf6b71 Merge pull request #5080 from jeremymcs/fix/headless-auto-cwd-anchor
+```
+
+Effort: 2-3 hours. Touches worktree code we already heavily customized — **conflicts likely**.
+
+---
+
+## F. Workflow state machine hardening (5 commits)
+
+```
+f2377eedd fix(auto): harden workflow state transitions
+b9a1c6743 fix(auto): persist workflow retry and summary state
+153fb328a fix(auto): address peer review state hardening
+381ccdef5 fix(state): fail closed on unreadable milestone summaries
+371b2eb31 fix(state): restore slice dependency fallback
+71e2c4b8d test(state): align dependency fallback expectation
+767c235fa Merge pull request #4758 from jeremymcs/fix/workflow-state-machine-hardening
+```
+
+Effort: 1 hour. Important for reliability of long auto runs.
+
+---
+
+## G. Provider additions (4 commits)
+
+Non-controversial provider list updates.
+
+```
+838dbc9b7 feat(models): add GLM-5.1 to Z.AI provider in custom models
+b21f936ce feat(models): add gpt-5.4-mini to openai-codex list (#1215)
+ba06f35c3 feat(gsd): add GPT-5.5 Codex model support
+5f3c90bd2 feat(ollama): native /api/chat provider with full option exposure
+6132d4089 feat(ollama): configurable probe/request timeouts via env vars
+939b75e45 Merge pull request #5045 from jeremymcs/feat/5003-ollama-timeout-env
+```
+
+Effort: <30 min. Mostly config/data.
+
+---
+
+## H. Security / data-integrity fixes (~6 commits)
+
+```
+65ca5aa2e fix(security): harden project-controlled surfaces  # we have 66ff949c1 partial; supersede
+da7dd56e7 fix(safety): persist bash evidence at tool_call to close mid-unit re-dispatch race (#5056)
+4370bedf3 fix(search): narrow native web_search injection to providers that accept it
+9340f1e9b fix(gsd): self-heal symlinked .gsd staging to prevent silent data loss (#4423)
+58d3d4d6c fix(knowledge): scope + budget milestone KNOWLEDGE injection (#4721)
+bb747ec57 fix(mcp-server): prevent defaultExecFn stdout-buffer deadlock
+```
+
+Effort: 1-2 hours. Most are surgical.
+
+---
+
+## I. Headless / non-interactive (5 commits)
+
+```
+4ba746888 fix(gsd): instruct workflows to use repo MCP tools
+14ec4d97f fix(headless): suppress notification status spam
+42f44f1ed fix(gsd): load global mcp and search providers
+c15afb45f fix(headless): improve search and mcp status output
+cf0274c63 fix(headless): show assistant previews in logs
+```
+
+Effort: 1 hour. Useful for our non-interactive autopilot path.
+
+---
+
+## J. Rate limiting + token telemetry (5 commits)
+
+```
+f980929f1 feat(auto): proactive rate limiting via min_request_interval_ms (#2996)
+73bc4d2f1 fix(auto): stamp request interval at dispatch
+41edad041 Merge pull request #5007 from jeremymcs/feat/min-request-interval-ms
+b4d4725ad feat(pi-coding-agent): opt-in per-call token telemetry (#5023)
+a400838aa Merge pull request #5026 from jeremymcs/feat/5023-token-telemetry
+```
+
+Effort: 1 hour. Aligns with SPEC.md §19.6 rate-limit observability.
+
+---
+
+## K. MCP global config (3 commits)
+
+```
+a59c38822 feat(mcp-client): read global MCP config from ~/.gsd/mcp.json
+49723ef03 Merge pull request #4970 from imxv/feat/mcp-client-global-config
+bb747ec57 fix(mcp-server): prevent defaultExecFn stdout-buffer deadlock
+```
+
+Effort: <1 hour.
+
+---
+
+## L. Doctor / diagnostics (2 commits)
+
+```
+420354f99 feat(gsd): add doctor check for orphan milestone directories (#4996)
+1fb9f439e Merge pull request #4998 from gsd-build/fix/4996-milestone-id-gap-detection
+```
+
+Effort: <30 min.
+
+---
+
+## M. Performance (3 commits)
+
+```
+4dd01472a Merge pull request #5030 from jeremymcs/perf/5027-compaction-cache-breakpoint
+8ebb13ee9 Merge pull request #5029 from jeremymcs/perf/5022-startup-optimization
+```
+
+Effort: <30 min if conflicts are minimal.
+
+---
+
+## N. Windows fixes (2 commits)
+
+```
+9d08d820b Merge pull request #5036 from TommyC81/fix/5015-windows-home-dir
+780a8220a Merge pull request #5042 from jeremymcs/fix/5017-windows-dep0190
+f857a68ba Merge pull request #5043 from jeremymcs/fix/4946-types-semver
+```
+
+Effort: <30 min. Take if Windows is a target; skip otherwise.
+
+---
+
+## O. UnitContextManifest / Composer rewrite (~15 commits)
+
+A major architectural refactor. **Likely conflicts heavily** with our work. Probably **skip** unless we want this direction; revisit during v3 implementation.
+
+```
+7d54fe2d3 feat(auto): UnitContextManifest schema + data + CI guard — phase 1 of #4782
+ae5b4011e feat(auto): UnitContextManifest v2 contract — typed computed artifacts (#4924)
+896da7915 feat(auto): UnitContextManifest tools-policy field — declarative-only (#4934)
+7a63d5558 feat(gsd): runtime tools-policy enforcement for planning units (#4934)
+1433c5f8e feat(auto): compose reassess-roadmap context from manifest — #4782 phase 2
+8a0eee56a feat(auto): migrate run-uat through composer — #4782 phase 3 batch 1
+dc9e7a854 feat(auto): migrate research-milestone through composer — #4782 phase 3 batch 2
+1765a211c feat(auto): migrate complete-slice through composer — #4782 phase 3 batch 3
+17b74c5bf feat(auto): wire pipeline variant into dispatch — phase 2 of #4781
+298d63707 feat(auto): milestone scope classifier — phase 1 of #4781
+4b4ab00f4 feat(unit-manifest): introduce planning-dispatch mode for slice plan/complete
+```
+
+Effort: 1-2 days IF we take it. **Recommendation: defer; revisit when v3 §3 schema reconciliation lands.**
+
+---
+
+## P. Memories cutover (1 commit — relevant for v3 sm integration)
+
+```
+d3600f92f feat(gsd): cutover to memories table as single source of truth (ADR-013 step 6)
+1f8e77172 Merge pull request #5002 from jeremymcs/fix/4967-memory-capture-error
+```
+
+Worth reading carefully — this is upstream's answer to what we're calling Singularity Memory integration. May change the recommended sm integration path in BUILD_PLAN.
+
+---
+
+## Recommended order of cherry-picks
+
+Total estimated effort if we take all clusters A–N: **~10-15 hours of focused work**, plus conflict resolution.
+
+| Order | Cluster | Why first |
+|---|---|---|
+| 1 | B agent-session | Critical correctness, lands cleanly |
+| 2 | F workflow state | Reliability of long auto runs |
+| 3 | H security/data-integrity | We already partially cherry-picked H#1 |
+| 4 | C claude-code permission | Small, isolated |
+| 5 | A eval-review | New feature, atomic block |
+| 6 | G providers | Trivial config |
+| 7 | J rate limiting | Aligns with §19.6 |
+| 8 | E worktree path safety | Conflicts likely; resolve carefully |
+| 9 | I headless | Useful for autopilot |
+| 10 | K MCP global config | Small |
+| 11 | L doctor / orphan check | Small |
+| 12 | D worktree TUI commands | Discretionary feature |
+| 13 | M performance | If gains are real |
+| 14 | N Windows | Skip if not a target |
+| **DEFER** | O composer rewrite | Conflicts; revisit during v3 |
+| **READ FIRST** | P memories cutover | Informs sm integration plan |
+
+## Excluded from this list
+
+- ~3,800 commits that are: chore, docs, test housekeeping, internal renames, CI tweaks, version bumps, dependency updates without our use case, branch-merge noise, revert-then-readd churn.
+- Most `Merge pull request` commits where the underlying squash already represents the work.
+
+If you want any of those clusters expanded with full file-touch lists before deciding, ask.