diff --git a/openspec/changes/rebuild-graph-without-blocking-writes/.openspec.yaml b/openspec/changes/rebuild-graph-without-blocking-writes/.openspec.yaml
new file mode 100644
index 00000000..8e7013b8
--- /dev/null
+++ b/openspec/changes/rebuild-graph-without-blocking-writes/.openspec.yaml
@@ -0,0 +1,2 @@
+schema: spec-driven
+created: 2026-07-27
diff --git a/openspec/changes/rebuild-graph-without-blocking-writes/design.md b/openspec/changes/rebuild-graph-without-blocking-writes/design.md
new file mode 100644
index 00000000..a8f7a1f2
--- /dev/null
+++ b/openspec/changes/rebuild-graph-without-blocking-writes/design.md
@@ -0,0 +1,105 @@
+## Context
+
+`EpistemicGraphIndex.refresh_paths` takes the vault mutation boundary and calls
+`_refresh_paths_locked`, which escalates to `_rebuild_all_locked()` whenever
+`available()` is false. `_rebuild_all_locked` runs a stabilization loop:
+snapshot `_disk_vault_freshness`, run `_rebuild_all_pass`, and accept the result
+only if freshness is unchanged; otherwise retry, up to
+`REBUILD_STABILIZATION_ATTEMPTS`, then fail and mark the graph unavailable.
+
+Two properties of the current implementation matter here.
+
+`_rebuild_all_pass` mutates the **live** sidecar: it deletes every row from
+`graph_edges`, `graph_nodes`, and `graph_parent_refs` before refilling. A reader
+arriving mid-pass would see an empty or partial graph. That is safe today only
+because the boundary excludes everyone for the whole rebuild.
+
+The stabilization loop is therefore currently a guard against *out-of-band*
+edits — a user editing files directly in Obsidian — rather than against
+concurrent Exomem writes, which the boundary already prevents.
+
+Moving the rebuild outside the boundary changes the status of both: the wipe
+becomes visible, and the stabilization loop becomes the primary consistency
+mechanism rather than a backstop.
+
+## Goals / Non-Goals
+
+**Goals:**
+
+- A full rebuild must not block unrelated vault mutations for its duration.
+- A reader must never observe a partially rebuilt graph.
+- Preserve the write-path contract: a write against an unusable sidecar returns
+  with the graph built and available.
+- Preserve the failure contract: a rebuild that cannot observe a stable vault
+  marks the graph unavailable rather than publishing a bad one.
+
+**Non-Goals:**
+
+- Changing the sidecar schema or file format.
+- Making incremental (`available()`-true) refresh faster.
+- Removing the escalation itself. #346 established that it is load-bearing.
+- Reducing how often the sidecar is invalidated. Fewer `SCHEMA_VERSION` bumps or
+  a registry-hash change that does not invalidate would both reduce exposure,
+  but they are separate questions from how a rebuild behaves.
+
+## Decisions
+
+Build into a temporary database in the sidecar's own directory, then swap. Same
+directory so the publish step is an atomic rename on one filesystem. The rebuild
+passes run against the temp database with no boundary held; only the swap takes
+it. The boundary hold drops from the length of a full rebuild — 32 s at 2,000
+pages, 172 s at 8,000 — to a rename.
+
+Re-verify freshness under the boundary immediately before the swap. Outside the
+boundary the vault can change during the final pass, so the pre-swap check is
+what makes the published graph trustworthy. If freshness moved, the rebuild
+retries within its existing attempt budget rather than publishing.
+
+Make rebuilds single-flight per vault. Today the boundary serialises them
+implicitly: the second writer blocks, and by the time it runs the graph is
+available so it takes the incremental path. Once rebuilds no longer hold the
+boundary, N concurrent writers would otherwise each start a full rebuild. A
+rebuild-in-progress marker keeps one running and lets the others wait for its
+result.
+
+Keep `_mark_unavailable()` on stabilization failure. The failure contract is
+unchanged; only the location of the work moves.
+
+Sweep abandoned temp databases in `reconcile`. A crash mid-rebuild leaves a temp
+file that no longer has an owner. Reconcile already walks sidecar state and
+already owns `rebuild_all`, so it is the natural place. Temp databases use a
+reserved name prefix so a sweep cannot mistake one for the live sidecar.
+
+## Risks / Trade-offs
+
+- [The stabilization budget may be too small outside the boundary] → Concurrent
+  Exomem writes can now perturb freshness mid-pass, where previously only
+  out-of-band edits could. If `REBUILD_STABILIZATION_ATTEMPTS` proves
+  insufficient under sustained writes, a rebuild that used to succeed slowly
+  would fail instead. Covered by a test that writes continuously during a
+  rebuild; the attempt budget may need raising with the sweep.
+- [Disk usage doubles transiently during a rebuild] → One extra sidecar-sized
+  file for the duration. Acceptable, and bounded by the sweep.
+- [A waiting writer still waits] → Single-flight means a concurrent writer that
+  needs the graph still waits for the in-flight rebuild. It no longer blocks
+  *unrelated* mutations, which is the reported harm, but the first writer after
+  invalidation still pays the rebuild cost. Reducing that is a separate problem.
+- [Atomic rename semantics on Windows] → Replacing an open SQLite file differs
+  from POSIX. The swap must tolerate readers holding the old file, and the
+  implementation must verify behaviour on the Windows service path rather than
+  assuming POSIX rename.
+
+## Migration Plan
+
+No data migration. The first rebuild after this change writes a temp database
+and swaps it; existing sidecars are read normally until invalidated. Roll back
+by restoring the in-boundary rebuild, which changes no on-disk format.
+
+## Open Questions
+
+- What should a writer do while a rebuild is in flight — block on the
+  single-flight result, or return deferred and let the rebuild cover its path?
+  Deferring keeps the write bounded but means the writer's own edit may not be
+  in the published graph, which the current contract does not allow.
+- Does `REBUILD_STABILIZATION_ATTEMPTS` need to rise now that concurrent Exomem
+  writes can perturb freshness, and what is the right ceiling before failing?
diff --git a/openspec/changes/rebuild-graph-without-blocking-writes/proposal.md b/openspec/changes/rebuild-graph-without-blocking-writes/proposal.md
new file mode 100644
index 00000000..ddd3ff67
--- /dev/null
+++ b/openspec/changes/rebuild-graph-without-blocking-writes/proposal.md
@@ -0,0 +1,65 @@
+## Why
+
+A full epistemic graph rebuild runs inside the vault mutation boundary, so it
+blocks every other vault mutation for its entire duration. Measured on a
+maintainer workstation against synthetic vaults, on a single `upsert_after_write`
+with no usable sidecar:
+
+| vault | first write | steady-state write | ratio |
+|---|---|---|---|
+| 500 pages | 7,728 ms | 315 ms | 24.6x |
+| 2,000 pages | 32,382 ms | 1,186 ms | 27.3x |
+
+CI's write-latency benchmark records the same shape at larger scale:
+`hold_ms=39092` over 2,000 pages and `hold_ms=172205` over 8,000.
+
+This is not an edge case. `_open_read_snapshot` treats the sidecar as unusable
+whenever `schema_version`, `core_registry_version`, or `extension_registry_hash`
+disagrees with the running build, so a full rebuild is triggered by:
+
+- any release that bumps `SCHEMA_VERSION` — at 7 today, bumped seven times;
+- any change to the relation registry, including a user adding or editing one
+  extension relation;
+- a sidecar that is missing, deleted, or corrupt.
+
+The first write after any of those stalls for tens of seconds on a
+maintainer-sized vault, and every concurrent writer is blocked behind it. A
+client that gives up meanwhile sees a timeout on a write that then lands.
+
+An earlier attempt (#346) removed the escalation so writes deferred to
+`reconcile`. That was wrong and CI rejected it:
+`test_refresh_missing_sidecar_routes_to_full_rebuild` requires a write against a
+missing sidecar to index the whole vault and leave the graph available. The
+rebuild has to happen; it must stop holding the boundary while it does.
+
+## What Changes
+
+- Build a full graph rebuild into a temporary database outside the vault
+  mutation boundary, then acquire the boundary only to swap it into place.
+- Make concurrent rebuild requests single-flight, so N blocked writers do not
+  each start their own full-vault rebuild.
+- Keep the existing stabilization contract: a rebuild that cannot observe a
+  stable vault still fails and marks the graph unavailable.
+- Preserve the write-path contract that a write against an unusable sidecar
+  leaves the graph built and available.
+- Sweep abandoned temporary rebuild databases during reconcile.
+
+## Capabilities
+
+### New Capabilities
+
+- `graph-rebuild-availability`: A full epistemic graph rebuild does not block
+  unrelated vault mutations, and never exposes a partially rebuilt graph.
+
+### Modified Capabilities
+
+None.
+
+## Impact
+
+Affected areas are `src/exomem/epistemic_graph.py`, `src/exomem/reconcile.py`
+(temp sweep), and the epistemic graph freshness, boundary, and semantic unit
+graph tests.
+
+No MCP tool schema, vault format, OAuth, or stdio behavior changes. The sidecar
+file format is unchanged; only how and where it is produced changes.
diff --git a/openspec/changes/rebuild-graph-without-blocking-writes/specs/graph-rebuild-availability/spec.md b/openspec/changes/rebuild-graph-without-blocking-writes/specs/graph-rebuild-availability/spec.md
new file mode 100644
index 00000000..8f8b1314
--- /dev/null
+++ b/openspec/changes/rebuild-graph-without-blocking-writes/specs/graph-rebuild-availability/spec.md
@@ -0,0 +1,58 @@
+## ADDED Requirements
+
+### Requirement: A full rebuild does not hold the vault mutation boundary
+The system SHALL perform full epistemic graph rebuild work without holding the
+vault mutation boundary, acquiring it only to publish the result.
+
+#### Scenario: Unrelated mutation during a rebuild
+- **WHEN** a full graph rebuild is in progress
+- **AND** an unrelated vault mutation is requested
+- **THEN** that mutation proceeds without waiting for the rebuild to finish
+
+#### Scenario: Boundary hold is bounded by the swap
+- **WHEN** a full graph rebuild completes
+- **THEN** the vault mutation boundary is held only for the publish step
+- **AND** the hold does not scale with vault size
+
+### Requirement: A partially rebuilt graph is never observable
+The system SHALL publish a rebuilt graph as a single atomic replacement.
+
+#### Scenario: Read during a rebuild
+- **WHEN** a graph read occurs while a full rebuild is in progress
+- **THEN** it observes either the previous graph state or the fully rebuilt one
+- **AND** never an empty or partially populated graph
+
+#### Scenario: Crash during a rebuild
+- **WHEN** the process terminates during a full rebuild
+- **THEN** the live sidecar is left in its pre-rebuild state
+- **AND** the abandoned temporary database is removed by the next reconcile
+
+### Requirement: Concurrent rebuild requests are single-flight
+The system SHALL run at most one full graph rebuild per vault at a time.
+
+#### Scenario: Several writers arrive with an unusable sidecar
+- **WHEN** multiple writes require a rebuild concurrently
+- **THEN** exactly one full rebuild runs
+- **AND** the remaining requests resolve from that rebuild's outcome
+
+### Requirement: Published graphs reflect a stable vault
+The system SHALL verify vault freshness immediately before publishing a rebuilt
+graph, while holding the vault mutation boundary.
+
+#### Scenario: Vault changes during the final pass
+- **WHEN** vault freshness changes between the start of a rebuild pass and the publish step
+- **THEN** the rebuild is retried rather than published
+
+#### Scenario: Rebuild cannot observe a stable vault
+- **WHEN** a rebuild exhausts its stabilization attempts
+- **THEN** it does not publish a graph
+- **AND** the graph is marked unavailable
+
+### Requirement: The write path still yields an available graph
+The system SHALL preserve the existing write-path contract when the sidecar is
+unusable.
+
+#### Scenario: Write against a missing sidecar
+- **WHEN** a write occurs and the graph sidecar is missing, schema-mismatched, or registry-invalidated
+- **THEN** the graph is rebuilt across the whole vault
+- **AND** the graph reports itself available once the write returns
diff --git a/openspec/changes/rebuild-graph-without-blocking-writes/tasks.md b/openspec/changes/rebuild-graph-without-blocking-writes/tasks.md
new file mode 100644
index 00000000..9ce05e2c
--- /dev/null
+++ b/openspec/changes/rebuild-graph-without-blocking-writes/tasks.md
@@ -0,0 +1,70 @@
+## 1. Regression coverage
+
+- [ ] 1.1 Prove an unrelated mutation is not blocked while a full rebuild runs.
+- [ ] 1.2 Prove a reader during a rebuild sees the old graph or the new one, never a partial one.
+- [ ] 1.3 Prove concurrent rebuild requests run exactly one rebuild.
+- [ ] 1.4 Prove the existing write-path contract still holds: a write against an
+      unusable sidecar indexes the whole vault and leaves it available
+      (the contract #346 broke).
+- [ ] 1.5 Prove a vault change during the final pass retries rather than publishes.
+- [ ] 1.6 Prove exhausted stabilization still marks the graph unavailable.
+
+## 2. Rebuild off the boundary
+
+- [ ] 2.1 Build rebuild passes into a temporary database beside the sidecar.
+- [ ] 2.2 Publish by atomic replacement under a boundary hold scoped to the swap.
+- [ ] 2.3 Re-verify freshness under the boundary immediately before publishing.
+- [ ] 2.4 Verify the replacement behaves correctly on the Windows service path,
+      where replacing an open SQLite file does not follow POSIX rename semantics.
+
+## 3. Single-flight
+
+- [ ] 3.1 Serialize rebuilds per vault so concurrent writers do not each start one.
+- [ ] 3.2 Resolve waiting requests from the in-flight rebuild's outcome.
+- [ ] 3.3 Settle the open question: whether a waiting writer blocks or defers,
+      and reconcile that with the contract in 1.4.
+
+## 4. Housekeeping
+
+- [ ] 4.1 Name temporary rebuild databases with a reserved prefix.
+- [ ] 4.2 Sweep abandoned temporary databases during reconcile.
+
+## 5. Verification
+
+- [ ] 5.1 Run the epistemic graph, freshness, boundary, semantic unit graph, and
+      reconcile suites, plus Ruff and strict OpenSpec validation.
+- [ ] 5.2 Re-run the reproduction harness and record the new first-write cost at
+      500 and 2,000 pages against the pre-change baseline.
+- [ ] 5.3 Confirm under a concurrent-write load that the stabilization budget is
+      still sufficient, and raise `REBUILD_STABILIZATION_ATTEMPTS` if not.
+- [ ] 5.4 Run the lean suite.
+
+### Baseline to beat
+
+Measured on a maintainer workstation against synthetic dense vaults, on main,
+one `upsert_after_write` with no usable sidecar:
+
+| vault | first write | steady-state write | ratio |
+|---|---|---|---|
+| 500 pages | 7,727.8 ms | 314.5 ms | 24.6x |
+| 2,000 pages | 32,381.5 ms | 1,186.2 ms | 27.3x |
+
+with `vault mutation boundary held too long ...
+operation=epistemic_graph_refresh_paths holder_kind=graph hold_ms=32371.03`.
+
+CI's write-latency benchmark shows `hold_ms=39092` at 2,000 pages and
+`hold_ms=172205` at 8,000.
+
+The target is that the *boundary hold* stops scaling with vault size. The first
+writer after invalidation still pays the rebuild; what must stop is every other
+mutation waiting behind it.
+
+### Prior art
+
+#346 attempted to fix this by removing the escalation so writes deferred to
+reconcile. CI rejected it:
+`test_refresh_missing_sidecar_routes_to_full_rebuild` requires a write against a
+missing sidecar to index the whole vault and leave it available, and three other
+tests depend on the same escalation — the spawned-mutator boundary contract,
+forced re-resolution after a relation-registry change, and the schema-v3 sidecar
+migration. Task 1.4 exists to keep that contract pinned while the rebuild moves.
diff --git a/scripts/repro_graph_rebuild_hold.py b/scripts/repro_graph_rebuild_hold.py
new file mode 100644
index 00000000..6ebbb224
--- /dev/null
+++ b/scripts/repro_graph_rebuild_hold.py
@@ -0,0 +1,67 @@
+"""Measure the mutation-boundary hold caused by one write against a missing graph sidecar.
+
+Reproduces on main's code path (no fix applied). Reports, per vault size:
+  - how long a SINGLE `upsert_after_write` takes when the sidecar is absent
+  - how long the same write takes once the sidecar exists (the healthy case)
+
+The gap between the two is the stall a user eats on the first write after the
+sidecar is missing, deleted, schema-bumped, or registry-invalidated.
+"""
+
+from __future__ import annotations
+
+import os
+import sys
+import tempfile
+import time
+from pathlib import Path
+
+ROOT = Path(__file__).resolve()
+REPO = Path(sys.argv[1]) if len(sys.argv) > 1 else None
+if REPO is None:
+    raise SystemExit("usage: repro_graph_hold.py <repo-root> [sizes...]")
+sys.path.insert(0, str(REPO / "src"))
+sys.path.insert(0, str(REPO / "scripts"))
+
+for name in (
+    "EXOMEM_DISABLE_EMBEDDINGS",
+    "EXOMEM_DISABLE_CLIP",
+    "EXOMEM_DISABLE_MEDIA_EXTRACTION",
+    "EXOMEM_DISABLE_RANKING",
+):
+    os.environ[name] = "1"
+
+from synth_vault import gen_dense_vault  # noqa: E402
+
+from exomem import epistemic_graph  # noqa: E402
+from exomem.kbdir import kb_dirname  # noqa: E402
+
+SIZES = [int(a) for a in sys.argv[2:]] or [500, 2000]
+
+
+def one_write(vault: Path, target: Path) -> float:
+    started = time.perf_counter()
+    epistemic_graph.upsert_after_write(vault, [target])
+    return (time.perf_counter() - started) * 1000.0
+
+
+for size in SIZES:
+    with tempfile.TemporaryDirectory(prefix=f"graph-hold-{size}-") as temp:
+        vault = Path(temp) / "vault"
+        vault.mkdir(parents=True)
+        gen_dense_vault(vault, size, links_per_note=3)
+
+        target = next((vault / kb_dirname()).rglob("*.md"))
+        sidecar = epistemic_graph.sidecar_path(vault)
+        assert not sidecar.exists(), "precondition: no sidecar yet"
+
+        cold_ms = one_write(vault, target)
+        built = sidecar.exists()
+        warm_ms = one_write(vault, target)
+
+        print(
+            f"pages={size:<6} first_write_missing_sidecar={cold_ms:9.1f}ms  "
+            f"subsequent_write={warm_ms:7.1f}ms  "
+            f"ratio={cold_ms / max(warm_ms, 0.001):7.1f}x  sidecar_built={built}",
+            flush=True,
+        )