diff --git a/CLAUDE.md b/CLAUDE.md
index 8659a11f1e..fab3f2672f 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -50,14 +50,14 @@ TechEmpower's production fork of [MemPalace/mempalace](https://github.com/MemPal
 
 - `~/Projects/mempalace.yaml` — **do not delete**. Mining config with wing/room definitions. Regenerate with `mempalace init ~/Projects --yes` if lost.
 - `~/.mempalace/config.json` — topic wings and hall keywords, customized for JP's domains (infrastructure, development, tools, creative, projects, system).
-- `~/.mempalace/palace/` — ChromaDB vector store. The actual data.
+- **Palace data lives in postgres on `disks.jphe.in`**, reached via palace-daemon at `http://disks.jphe.in:8085`. The local `~/.mempalace/palace/` ChromaDB store was retired on the 2026-05-14 pgvector cutover (renamed to `~/.mempalace/palace.retired-pre-pgcutover-2026-05-14/`); a `~/.mempalace/RETIRED` marker refuses default-palace opens so misconfigured tools can't silently fall back to the stale local store.
 - `~/.mempalace/hook_state/` — stop hook session tracking.
 
 ## Development
 
 ```bash
-source venv/bin/activate
-python -m pytest tests/ -q              # ~1096 tests (benchmarks deselected)
+source .venv/bin/activate
+python -m pytest tests/ -q              # 2486 tests (benchmarks deselected)
 mempalace status                         # check palace state
 mempalace search "query"                 # test search
 python -m mempalace.mcp_server           # run MCP server standalone
@@ -83,7 +83,7 @@ Workflow for landing new fork-ahead changes lives in [Documentation maintenance]
 Claude Code has two complementary memory layers, used in tandem:
 
 - **Auto-memory** (`~/.claude/projects/*/memory/`) — lightweight preferences, context, feedback. Manual writes only. (Anthropic's "Auto Dream" research-preview shipped late April 2026 in Claude Code `/dream` + the Managed Agents Dreams API; MemPalace deliberately stays un-consolidated and the Dreams API design ratifies the verbatim-vs-derivative axis. See `~/.claude/projects/-home-jp-Projects-memorypalace/memory/project_auto_dream.md`.)
-- **MemPalace** (`~/.mempalace/palace/`, ~183K drawers behind the daemon) — verbatim conversations, tool output, code. Write-only archive, searchable via MCP. Completeness is the feature. Wing/room layout follows the canonical 7-room taxonomy (see `palace-taxonomy` skill / `~/Projects/familiar.realm.watch/docs/superpowers/specs/2026-05-13-palace-room-taxonomy.md`).
+- **MemPalace** (postgres + pgvector + AGE on `disks.jphe.in`, reached via palace-daemon at `http://disks.jphe.in:8085`, 300K+ drawers) — verbatim conversations, tool output, code. Write-only archive, searchable via MCP. Completeness is the feature. Wing/room layout follows the canonical 7-room taxonomy (see `palace-taxonomy` skill / `~/Projects/familiar.realm.watch/docs/superpowers/specs/2026-05-13-palace-room-taxonomy.md`).
 
 Both systems coexist. Hook saves are scoped to MemPalace ("For THIS save, use MemPalace MCP tools only") — this is not a permanent ban on auto-memory.
 
diff --git a/FORK_CHANGELOG.md b/FORK_CHANGELOG.md
index 4e9e9c80e0..e88658d90e 100644
--- a/FORK_CHANGELOG.md
+++ b/FORK_CHANGELOG.md
@@ -1,4 +1,4 @@
-# Fork Changelog (jphein/mempalace)
+# Fork Changelog (techempower-org/mempalace)
 
 Fork-ahead changes that aren't yet in upstream `MemPalace/mempalace`.
 Upstream's release history lives in [`CHANGELOG.md`](CHANGELOG.md);
@@ -24,7 +24,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 ### Added
 
 
-- **Honor ~/.mempalace/RETIRED marker — refuse default palace, surface retire message** ([`798cf14`](https://github.com/jphein/mempalace/commit/798cf14))
+- **Honor ~/.mempalace/RETIRED marker — refuse default palace, surface retire message** ([`798cf14`](https://github.com/techempower-org/mempalace/commit/798cf14))
   Recurring confusion source on this fork: any code path that opens
   mempalace without ``PALACE_DAEMON_URL`` set silently falls
   through to the default chroma palace at ``~/.mempalace/palace``
@@ -66,7 +66,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 ### Fixed
 
 
-- **Empty repo .opencode/opencode.json mcp block — disabled flag wasn't being respected** ([`7133eee`](https://github.com/jphein/mempalace/commit/7133eee))
+- **Empty repo .opencode/opencode.json mcp block — disabled flag wasn't being respected** ([`7133eee`](https://github.com/techempower-org/mempalace/commit/7133eee))
   ``.opencode/opencode.json`` from upstream PR #1567 spawns a
   local ``mempalace.mcp_server`` when opencode launches in the
   repo root. On daemon-routed setups this creates **two MCP
@@ -85,7 +85,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
   *Files:* `.opencode/opencode.json`
 
 
-- **Drop \$comment from .opencode/opencode.json — schema rejects unknown root keys** ([`637bb01`](https://github.com/jphein/mempalace/commit/637bb01))
+- **Drop \$comment from .opencode/opencode.json — schema rejects unknown root keys** ([`637bb01`](https://github.com/techempower-org/mempalace/commit/637bb01))
   Commit ``47018e5`` (PR #108) added a ``\$comment`` field to
   ``.opencode/opencode.json`` explaining why the repo-level MCP
   entry defaulted to ``enabled: false``. Opencode's strict
@@ -104,7 +104,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
   *Files:* `.opencode/opencode.json`
 
 
-- **Disable repo-level MCP entry by default + venv-python fallback** ([`47018e5`](https://github.com/jphein/mempalace/commit/47018e5))
+- **Disable repo-level MCP entry by default + venv-python fallback** ([`47018e5`](https://github.com/techempower-org/mempalace/commit/47018e5))
   First attempt at the daemon-routed-vs-local-spawn double-server
   problem from upstream PR #1567's ``.opencode/opencode.json``.
   Set ``enabled: false`` on the repo entry; switched the spawn
@@ -127,7 +127,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 ### Added
 
 
-- **Bundled OpenCode live-capture plugin that bypasses option-K v1.2.1 bugs (filed upstream as #4, #5)** ([`5522623`](https://github.com/jphein/mempalace/commit/5522623))
+- **Bundled OpenCode live-capture plugin that bypasses option-K v1.2.1 bugs (filed upstream as #4, #5)** ([`5522623`](https://github.com/techempower-org/mempalace/commit/5522623))
   Adds `examples/opencode/live-capture/` — a self-contained
   OpenCode plugin (JS) + Python helper that POSTs verbatim
   session transcripts to the daemon's `/silent-save` endpoint on
@@ -182,7 +182,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
   *Files:* `examples/opencode/live-capture/mempalace-live-capture.js`, `examples/opencode/live-capture/capture-session.py`, `examples/opencode/option-k-plugin-daemon-routing.patch`, `examples/opencode/option-k-plugin-message-updated.patch`, `docs/integrations/opencode.md`
 
 
-- **Documented OpenCode integration recipe (read-side MCP + push plugin + retrospective adapter)** ([`60dc9e6`](https://github.com/jphein/mempalace/commit/60dc9e6))
+- **Documented OpenCode integration recipe (read-side MCP + push plugin + retrospective adapter)** ([`60dc9e6`](https://github.com/techempower-org/mempalace/commit/60dc9e6))
   Adds `docs/integrations/opencode.md` and an `examples/opencode/`
   directory capturing the three-direction OpenCode + MemPalace
   integration recipe for daemon-routed setups:
@@ -222,7 +222,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
   *Files:* `docs/integrations/opencode.md`, `examples/opencode/opencode.jsonc.example`, `examples/opencode/option-k-plugin-daemon-routing.patch`
 
 
-- **.opencode/opencode.json — repo-root MCP config so opencode picks up mempalace automatically** ([`ba16b82`](https://github.com/jphein/mempalace/commit/ba16b82))
+- **.opencode/opencode.json — repo-root MCP config so opencode picks up mempalace automatically** ([`ba16b82`](https://github.com/techempower-org/mempalace/commit/ba16b82))
   Cherry-pick of upstream PR #1567 (Dxrk777). Adds
   `.opencode/opencode.json` so that running `opencode` in the
   mempalace repo root automatically wires `mempalace` as a local
@@ -245,7 +245,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
   *Files:* `.opencode/opencode.json`
 
 
-- **OpenCodeSourceAdapter (RFC 002) — retrospective ingest of OpenCode SQLite sessions** ([`2ffe652`](https://github.com/jphein/mempalace/commit/2ffe652))
+- **OpenCodeSourceAdapter (RFC 002) — retrospective ingest of OpenCode SQLite sessions** ([`2ffe652`](https://github.com/techempower-org/mempalace/commit/2ffe652))
   Cherry-pick of upstream PR #1484. Adds
   `mempalace/sources/opencode.py` — an RFC 002 `BaseSourceAdapter`
   that ingests OpenCode AI-coding-CLI session transcripts from
@@ -284,7 +284,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
   *Files:* `mempalace/sources/opencode.py`, `mempalace/sources/transforms.py`, `mempalace/sources/context.py`, `pyproject.toml`, `tests/test_sources_opencode.py`, `tests/fixtures/opencode/sample_session_2026_05_12/README.md`, `tests/fixtures/opencode/sample_session_2026_05_12/build_fixture.py`, `tests/test_corpus_origin_integration.py`
 
 
-- **Pending-writes journal + replay so daemon outages stop being silent** ([`0c34464`](https://github.com/jphein/mempalace/commit/0c34464))
+- **Pending-writes journal + replay so daemon outages stop being silent** ([`0c34464`](https://github.com/techempower-org/mempalace/commit/0c34464))
   Closes a silent-data-loss gap exposed by the 2026-05-17 power
   event: when ``PALACE_DAEMON_URL`` is set and the daemon's
   backend is unreachable, the hook write path (``_post_daemon_mine``)
@@ -315,7 +315,7 @@ enqueue side effect.
 ### Fixed
 
 
-- **Stub resources/list + prompts/list so MCP clients stop ERROR-logging on connect** ([`6ca0670`](https://github.com/jphein/mempalace/commit/6ca0670))
+- **Stub resources/list + prompts/list so MCP clients stop ERROR-logging on connect** ([`6ca0670`](https://github.com/techempower-org/mempalace/commit/6ca0670))
   OpenCode 1.15.x and other MCP clients probe ``resources/list``
   and ``prompts/list`` on connect to discover server
   capabilities. mempalace's MCP server only exposes ``tools/*``;
@@ -334,7 +334,7 @@ empty-list shape and absence of ERROR.
   *Files:* `mempalace/mcp_server.py`, `tests/test_mcp_server.py`
 
 
-- **MCP server distinguishes 'backend unreachable' from 'no palace found'** ([`0c34464`](https://github.com/jphein/mempalace/commit/0c34464))
+- **MCP server distinguishes 'backend unreachable' from 'no palace found'** ([`0c34464`](https://github.com/techempower-org/mempalace/commit/0c34464))
   The CLI's misleading ``Run: mempalace init <dir>`` hint was
   the diagnostic blocker during the 2026-05-17 outage: the daemon
   was up, postgres was down, and every search returned "No palace
@@ -369,7 +369,7 @@ on failure, error cleared on successful reopen).
 ### Added
 
 
-- **mempalace_walk_palace MCP tool — agent walks the palace via AGE Cypher** ([`8022ecb`](https://github.com/jphein/mempalace/commit/8022ecb))
+- **mempalace_walk_palace MCP tool — agent walks the palace via AGE Cypher** ([`8022ecb`](https://github.com/techempower-org/mempalace/commit/8022ecb))
   Phase 6 of the AGE-integration plan. Exposes the "agent walks into
   the palace finding wings, rooms, drawers" metaphor as a single MCP
   tool over the unified palace+entity graph (Wing → Room → Drawer →
@@ -397,7 +397,7 @@ on failure, error cleared on successful reopen).
   *Files:* `mempalace/mcp_server.py`
 
 
-- **Backfill AGE graph from existing drawer table — restartable, checkpointed** ([`b3f0206`](https://github.com/jphein/mempalace/commit/b3f0206))
+- **Backfill AGE graph from existing drawer table — restartable, checkpointed** ([`b3f0206`](https://github.com/techempower-org/mempalace/commit/b3f0206))
   Phase 4 of the AGE-integration plan. New module
   `mempalace/backfill_age.py` with CLI entry point that reads the
   drawer table once and builds the full Wing/Room/Drawer/MENTIONS
@@ -431,7 +431,7 @@ on failure, error cleared on successful reopen).
   *Files:* `mempalace/backfill_age.py`, `mempalace/knowledge_graph_age.py`
 
 
-- **Wing/Room/Drawer hierarchy as native AGE nodes; Cypher MATCH walks palace structure** ([`ff583c0`](https://github.com/jphein/mempalace/commit/ff583c0))
+- **Wing/Room/Drawer hierarchy as native AGE nodes; Cypher MATCH walks palace structure** ([`ff583c0`](https://github.com/techempower-org/mempalace/commit/ff583c0))
   Phase 3 of the AGE-integration plan. Mirrors `mempalace.palace_graph`'s
   SQL-aggregation pattern into AGE so Cypher MATCH walks the palace
   structure natively — no SQL aggregation per query.
@@ -467,7 +467,7 @@ on failure, error cleared on successful reopen).
   *Files:* `mempalace/palace_graph_age.py`
 
 
-- **Write-through middleware on PostgresCollection — entities populate AGE on every drawer write** ([`3321d83`](https://github.com/jphein/mempalace/commit/3321d83))
+- **Write-through middleware on PostgresCollection — entities populate AGE on every drawer write** ([`3321d83`](https://github.com/techempower-org/mempalace/commit/3321d83))
   Phase 2 of the AGE-integration plan. Adds a write-through hook on
   `PostgresCollection.add`/`upsert` that extracts entities from the
   document and creates `(Drawer)-[:MENTIONS]->(Entity)` edges in
@@ -505,7 +505,7 @@ on failure, error cleared on successful reopen).
   *Files:* `mempalace/backends/postgres.py`, `mempalace/kg_writethrough.py`
 
 
-- **KnowledgeGraphAGE API parity with SQLite KG: add_entity, invalidate, query_entity, query_relationship, timeline, seed_from_entity_facts** ([`ff7187d`](https://github.com/jphein/mempalace/commit/ff7187d))
+- **KnowledgeGraphAGE API parity with SQLite KG: add_entity, invalidate, query_entity, query_relationship, timeline, seed_from_entity_facts** ([`ff7187d`](https://github.com/techempower-org/mempalace/commit/ff7187d))
   Phase 1 of the AGE-integration plan. Brings `KnowledgeGraphAGE` to
   API parity with `mempalace.knowledge_graph.KnowledgeGraph` (the
   SQLite backend). Previously only `add_triple`, `query_triples`,
@@ -555,7 +555,7 @@ on failure, error cleared on successful reopen).
 ### Added
 
 
-- **KnowledgeGraphAGE skeleton — Apache AGE graph bootstrap over psycopg2** ([`a3ee623`](https://github.com/jphein/mempalace/commit/a3ee623))
+- **KnowledgeGraphAGE skeleton — Apache AGE graph bootstrap over psycopg2** ([`a3ee623`](https://github.com/techempower-org/mempalace/commit/a3ee623))
   First commit toward the Apache AGE-backed knowledge graph layer
   that the migration plan calls for. Skeleton class
   `KnowledgeGraphAGE` in `mempalace/knowledge_graph_age.py` opens a
@@ -608,7 +608,7 @@ set against the homelab mempalace-db at disks.jphe.in:5433).
   *Files:* `mempalace/knowledge_graph_age.py`, `tests/test_knowledge_graph_age.py`
 
 
-- **CI: gate postgres-backend tests against a pgvector service container** ([`da0bdbb`](https://github.com/jphein/mempalace/commit/da0bdbb))
+- **CI: gate postgres-backend tests against a pgvector service container** ([`da0bdbb`](https://github.com/techempower-org/mempalace/commit/da0bdbb))
   Adds a `test-postgres` job to `.github/workflows/ci.yml` that
   runs in parallel with the existing `test-linux` / `test-windows`
   / `test-macos` / `lint` matrix. Service container is the public
@@ -640,7 +640,7 @@ set against the homelab mempalace-db at disks.jphe.in:5433).
   *Files:* `.github/workflows/ci.yml`
 
 
-- **PostgreSQL backend via #665 cherry-pick + fork-side adaptations + smoke tests** ([`5e90c72`](https://github.com/jphein/mempalace/commit/5e90c72))
+- **PostgreSQL backend via #665 cherry-pick + fork-side adaptations + smoke tests** ([`5e90c72`](https://github.com/techempower-org/mempalace/commit/5e90c72))
   Cherry-pick of skuznetsov's upstream PR
   [#665](https://github.com/MemPalace/mempalace/pull/665) — adds a
   PostgreSQL backend built on the merged #995 / RFC 001
@@ -724,7 +724,7 @@ regressions in non-postgres paths.
 ### Changed
 
 
-- **README pivots to the four-layer model + Auto Dream as vindication of the verbatim-vs-derivative axis** ([`55b36ca`](https://github.com/jphein/mempalace/commit/55b36ca))
+- **README pivots to the four-layer model + Auto Dream as vindication of the verbatim-vs-derivative axis** ([`55b36ca`](https://github.com/techempower-org/mempalace/commit/55b36ca))
   Substantial README rewrite (+137/-100) reflecting three things
   that landed between the previous refresh (`a67be3f`, the
   2026-05-10 develop sync) and now:
@@ -782,7 +782,7 @@ regressions in non-postgres paths.
 ### Fixed
 
 
-- **Defense-in-depth metadata sanitizer at the chromadb-client chokepoint** ([`f499814`](https://github.com/jphein/mempalace/commit/f499814))
+- **Defense-in-depth metadata sanitizer at the chromadb-client chokepoint** ([`f499814`](https://github.com/techempower-org/mempalace/commit/f499814))
   Companion to the repair.py sanitizers in #1458 / `949cb20`
   (which fixed `_extract_drawers` and `_rebuild_one_collection`).
   A 151,478-drawer rebuild against the canonical palace still
@@ -818,7 +818,7 @@ regressions in non-postgres paths.
   *Files:* `mempalace/backends/chroma.py`
 
 
-- **Coerce empty + None metadata to sentinel in both rebuild paths** ([`949cb20`](https://github.com/jphein/mempalace/commit/949cb20))
+- **Coerce empty + None metadata to sentinel in both rebuild paths** ([`949cb20`](https://github.com/techempower-org/mempalace/commit/949cb20))
   ChromaDB 1.5.x rejects both None and empty-dict entries in the
   `metadatas` list (raises `ValueError: Expected metadata to be a
   non-empty dict`). Two functions in `mempalace/repair.py` construct
@@ -855,7 +855,7 @@ regressions in non-postgres paths.
   *Files:* `mempalace/repair.py`
 
 
-- **Route Stop/PreCompact hooks through palace-daemon/clients/hook.py** ([`42ded2e`](https://github.com/jphein/mempalace/commit/42ded2e))
+- **Route Stop/PreCompact hooks through palace-daemon/clients/hook.py** ([`42ded2e`](https://github.com/techempower-org/mempalace/commit/42ded2e))
   Replaces the bash wrapper invocation pattern in
   `.claude-plugin/hooks/hooks.json` with a single Python entrypoint
   via the daemon's hook client. Both Stop and PreCompact now invoke
@@ -892,7 +892,7 @@ regressions in non-postgres paths.
 ### Performance
 
 
-- **Bulk pre-fetch already-mined set instead of N WHERE queries in mine_convos** ([`248854a`](https://github.com/jphein/mempalace/commit/248854a))
+- **Bulk pre-fetch already-mined set instead of N WHERE queries in mine_convos** ([`248854a`](https://github.com/techempower-org/mempalace/commit/248854a))
   Replaces the N+1 `col.get(where={"source_file": <path>}, ...)`
   per-conversation pattern in `mempalace/convo_miner.py:mine_convos`
   with a single bulk pre-fetch — `col.get(where={"source_file":
@@ -923,7 +923,7 @@ regressions in non-postgres paths.
 ### Added
 
 
-- **daemon-route `mempalace status` / `search` / `mine` when PALACE_DAEMON_URL is set** ([`22ef562`](https://github.com/jphein/mempalace/commit/22ef562))
+- **daemon-route `mempalace status` / `search` / `mine` when PALACE_DAEMON_URL is set** ([`22ef562`](https://github.com/techempower-org/mempalace/commit/22ef562))
   Companion to the `mcp_server` routing in commit `41359ba`. Closes
   the last desktop-side path that opened a local chromadb client.
 
@@ -969,7 +969,7 @@ when env var is unset. Suite 1591 passed (1577 + 14 new).
   *Files:* `mempalace/cli.py`, `tests/test_cli_daemon.py`
 
 
-- **daemon-route `mcp_server.py` via the `handle_request` JSON-RPC chokepoint** ([`41359ba`](https://github.com/jphein/mempalace/commit/41359ba))
+- **daemon-route `mcp_server.py` via the `handle_request` JSON-RPC chokepoint** ([`41359ba`](https://github.com/techempower-org/mempalace/commit/41359ba))
   Mirrors the `PALACE_DAEMON_URL` gate that `hooks_cli.py` shipped
   on 2026-04-24 (the daemon-strict fix for the HNSW drift
   incident). Closes the last in-process write path inside
@@ -1024,7 +1024,7 @@ from the canonical palace. Suite 1577 passed.
 ### Added
 
 
-- **mempalace mined + purge --source-file (mining management surface)** ([`2e6ced9`](https://github.com/jphein/mempalace/commit/2e6ced9))
+- **mempalace mined + purge --source-file (mining management surface)** ([`2e6ced9`](https://github.com/techempower-org/mempalace/commit/2e6ced9))
   Closes the "removing manually mined data" half of JP's
   mining-management ask. Adding is already covered by the existing
   ``mempalace mine <dir>``; this PR adds the symmetric remove +
@@ -1052,7 +1052,7 @@ from the canonical palace. Suite 1577 passed.
   *Files:* `mempalace/cli.py`, `tests/test_cli.py`
 
 
-- **`hook_verbatim_mode` config flag preserves system tags + full tool I/O during transcript ingest** ([`ef98961`](https://github.com/jphein/mempalace/commit/ef98961))
+- **`hook_verbatim_mode` config flag preserves system tags + full tool I/O during transcript ingest** ([`ef98961`](https://github.com/techempower-org/mempalace/commit/ef98961))
   `normalize()` defaults match upstream — system tags, hook chrome,
   Read/Edit/Write tool results, long Bash output, and large
   Grep/Glob match lists are stripped or truncated so chunk
@@ -1099,7 +1099,7 @@ config-file readback. Suite total 1562 passed.
 ### Changed
 
 
-- **Drop wing_ prefix from transcript-derived wings to converge with operator mines** ([`86d4700`](https://github.com/jphein/mempalace/commit/86d4700))
+- **Drop wing_ prefix from transcript-derived wings to converge with operator mines** ([`86d4700`](https://github.com/techempower-org/mempalace/commit/86d4700))
   The fork-only ``_wing_from_transcript_path`` returned
   ``wing_<project>`` for hook-derived wings, but operator-mined
   content from ``mempalace mine ~/Projects/X`` lands in a bare-name
@@ -1127,7 +1127,7 @@ config-file readback. Suite total 1562 passed.
   *Files:* `mempalace/hooks_cli.py`, `tests/test_hooks_cli.py`
 
 
-- **Retire mempalace_session_recovery collection + read tool** ([`0b945e1`](https://github.com/jphein/mempalace/commit/0b945e1))
+- **Retire mempalace_session_recovery collection + read tool** ([`0b945e1`](https://github.com/techempower-org/mempalace/commit/0b945e1))
   Follow-up to drop-checkpoint-write-path. With nothing writing
   to the recovery collection anymore (hooks moved to verbatim-only
   on the parent branch), the read paths and migration code that
@@ -1163,7 +1163,7 @@ config-file readback. Suite total 1562 passed.
   *Files:* `mempalace/palace.py`, `mempalace/mcp_server.py`, `mempalace/migrate.py`, `mempalace/cli.py`, `website/reference/mcp-tools.md`, `tests/test_session_recovery.py`, `tests/test_migrate.py`, `tests/test_mcp_server.py`
 
 
-- **Drop hook-side checkpoint diary writes — verbatim-only architecture** ([`69768fc`](https://github.com/jphein/mempalace/commit/69768fc))
+- **Drop hook-side checkpoint diary writes — verbatim-only architecture** ([`69768fc`](https://github.com/techempower-org/mempalace/commit/69768fc))
   The Stop hook used to do two things on each fire: (a) write a
   1KB checkpoint summary diary entry into the dedicated
   ``mempalace_session_recovery`` collection AND (b) auto-mine the
@@ -1206,7 +1206,7 @@ config-file readback. Suite total 1562 passed.
 ### Fixed
 
 
-- **Preserve dashed project names in transcript-derived wings** ([`d76134d`](https://github.com/jphein/mempalace/commit/d76134d))
+- **Preserve dashed project names in transcript-derived wings** ([`d76134d`](https://github.com/techempower-org/mempalace/commit/d76134d))
   Two findings from Copilot review on jphein/mempalace#9 that
   surfaced a real bug: the previous primary regex's
   ``encoded.rsplit('-', 1)[-1]`` rule collapsed
@@ -1232,7 +1232,7 @@ config-file readback. Suite total 1562 passed.
   *Files:* `mempalace/hooks_cli.py`, `tests/test_hooks_cli.py`
 
 
-- **Restore transcript ingest via daemon /mine when PALACE_DAEMON_URL is set** ([`09d2ca6`](https://github.com/jphein/mempalace/commit/09d2ca6))
+- **Restore transcript ingest via daemon /mine when PALACE_DAEMON_URL is set** ([`09d2ca6`](https://github.com/techempower-org/mempalace/commit/09d2ca6))
   Daemon-strict mode (introduced 2026-04-24 in commits ``8c90c0f``
   + ``0e97b19`` to fix the HNSW drift incident) skipped all three
   local mining paths when ``PALACE_DAEMON_URL`` was set, on the
@@ -1274,7 +1274,7 @@ config-file readback. Suite total 1562 passed.
 ### Fixed
 
 
-- **`cfg.init()` no longer materializes chunking defaults into `config.json`** ([`6ce37c0`](https://github.com/jphein/mempalace/commit/6ce37c0))
+- **`cfg.init()` no longer materializes chunking defaults into `config.json`** ([`6ce37c0`](https://github.com/techempower-org/mempalace/commit/6ce37c0))
   `cfg.init()` was unconditionally writing ``chunk_size: 800``,
   ``chunk_overlap: 100``, and ``min_chunk_size: 50`` into
   ``config.json`` on first run. The values match ``miner.py``'s
@@ -1312,7 +1312,7 @@ config-file readback. Suite total 1562 passed.
   merges. Amends fork-ahead row 17.
 
   *Tests:* 1548/1548 (was 1546/1548 with 2 isolation failures in test_convo_miner)
-  *Upstream:* [PR #1024](https://github.com/MemPalace/mempalace/pull/1024) (OPEN)
+  *Upstream:* [PR #1024](https://github.com/MemPalace/mempalace/pull/1024) (MERGED)
   *Files:* `mempalace/config.py`
 
 
@@ -1322,7 +1322,7 @@ config-file readback. Suite total 1562 passed.
 ### Changed
 
 
-- **Retire the `kind=` filter — structural split made it inert** ([`7ba28dc`](https://github.com/jphein/mempalace/commit/7ba28dc))
+- **Retire the `kind=` filter — structural split made it inert** ([`7ba28dc`](https://github.com/techempower-org/mempalace/commit/7ba28dc))
   Phases A–E of the checkpoint collection split (2026-04-25 → 2026-04-26)
   moved every Stop-hook auto-save checkpoint drawer to the dedicated
   ``mempalace_session_recovery`` collection. Empirical check on the
@@ -1343,7 +1343,7 @@ config-file readback. Suite total 1562 passed.
   *Files:* `mempalace/searcher.py`, `mempalace/mcp_server.py`, `mempalace/palace.py`, `mempalace/migrate.py`, `mempalace/layers.py`, `tests/test_searcher.py`
 
 
-- **Hoist CLOSET_RANK_BOOSTS to module level + record VecRecall ablation finding** ([`3cb03f3`](https://github.com/jphein/mempalace/commit/3cb03f3))
+- **Hoist CLOSET_RANK_BOOSTS to module level + record VecRecall ablation finding** ([`3cb03f3`](https://github.com/techempower-org/mempalace/commit/3cb03f3))
   Two-step refactor + measurement. First (commit ``f558d3c``):
   hoist ``CLOSET_RANK_BOOSTS = [0.40, 0.25, 0.15, 0.08, 0.04]`` and
   ``CLOSET_DISTANCE_CAP`` from inside ``search_memories`` to module
@@ -1372,7 +1372,7 @@ config-file readback. Suite total 1562 passed.
 ### Fixed
 
 
-- **Strip embedded API key from .claude-plugin/ manifests; rely on env inheritance** ([`9f91e18`](https://github.com/jphein/mempalace/commit/9f91e18))
+- **Strip embedded API key from .claude-plugin/ manifests; rely on env inheritance** ([`9f91e18`](https://github.com/techempower-org/mempalace/commit/9f91e18))
   ``.claude-plugin/.mcp.json`` and ``.claude-plugin/hooks/hooks.json``
   shipped with a real (rotated) API key embedded as a literal in the
   manifest's ``env`` block, plus my homelab daemon URL. Both are
@@ -1401,7 +1401,7 @@ config-file readback. Suite total 1562 passed.
 ### Added
 
 
-- **Canonical YAML manifest + renderer for fork-ahead docs** ([`5a01aec`](https://github.com/jphein/mempalace/commit/5a01aec))
+- **Canonical YAML manifest + renderer for fork-ahead docs** ([`5a01aec`](https://github.com/techempower-org/mempalace/commit/5a01aec))
   The fork-ahead narrative previously lived (and drifted) across four
   hand-edited files: README's fork-change-queue table, CLAUDE.md's row
   inventory, FORK_CHANGELOG.md, and the promises tracker. New
@@ -1418,7 +1418,7 @@ config-file readback. Suite total 1562 passed.
   *Files:* `docs/fork-changes.yaml`, `scripts/render-docs.py`, `scripts/check-docs.sh`, `FORK_CHANGELOG.md`, `CLAUDE.md`
 
 
-- **Phase D migration + PreCompact recovery write** ([`42817d7`](https://github.com/jphein/mempalace/commit/42817d7))
+- **Phase D migration + PreCompact recovery write** ([`42817d7`](https://github.com/techempower-org/mempalace/commit/42817d7))
   ``migrate_checkpoints_to_recovery(palace_path, batch_size=1000)`` walks
   the main collection in pages, filters drawers with topic in
   ``_CHECKPOINT_TOPICS`` in Python (avoids the chromadb 1.5.x ``$in``/``$nin``
@@ -1437,7 +1437,7 @@ config-file readback. Suite total 1562 passed.
   *Files:* `mempalace/migrate.py`, `mempalace/cli.py`, `mempalace/hooks_cli.py`, `tests/test_migrate.py`
 
 
-- **Surface drawer_id in search/diary/recovery payloads** ([`9a8bb77`](https://github.com/jphein/mempalace/commit/9a8bb77))
+- **Surface drawer_id in search/diary/recovery payloads** ([`9a8bb77`](https://github.com/techempower-org/mempalace/commit/9a8bb77))
   ChromaDB's primary key was always returned by ``query()`` and ``get()``
   but never plumbed into result-building loops; consumers (e.g.
   familiar.realm.watch's citation-popover loop) couldn't link a hit
@@ -1452,7 +1452,7 @@ config-file readback. Suite total 1562 passed.
   *Files:* `mempalace/searcher.py`, `mempalace/mcp_server.py`, `website/reference/mcp-tools.md`
 
 
-- **scripts/deploy.sh — one-command Syncthing-aware redeploy** ([`8252025`](https://github.com/jphein/mempalace/commit/8252025))
+- **scripts/deploy.sh — one-command Syncthing-aware redeploy** ([`8252025`](https://github.com/techempower-org/mempalace/commit/8252025))
   Single command does the right shape: push fork main → wait for
   Syncthing to reach ``/mnt/raid/projects/memorypalace`` on the deploy
   host → ``systemctl --user restart palace-daemon`` → poll ``/health`` →
@@ -1466,7 +1466,7 @@ config-file readback. Suite total 1562 passed.
 ### Changed
 
 
-- **Cherry-pick #1094 — coerce None metadatas at chromadb boundary** ([`43d728d`](https://github.com/jphein/mempalace/commit/43d728d))
+- **Cherry-pick #1094 — coerce None metadatas at chromadb boundary** ([`43d728d`](https://github.com/techempower-org/mempalace/commit/43d728d))
   Fork main was carrying the per-site ``meta = meta or {}`` guards
   from #999 in eight read paths but didn't have the boundary
   coercion that closes the issue once for all callers. The typed
@@ -1486,7 +1486,7 @@ config-file readback. Suite total 1562 passed.
   *Files:* `mempalace/backends/chroma.py`, `tests/test_backends.py`
 
 
-- **Cherry-pick #1087 rewrite — collection.delete(where=) instead of nuke-and-rebuild** ([`366a9ad`](https://github.com/jphein/mempalace/commit/366a9ad))
+- **Cherry-pick #1087 rewrite — collection.delete(where=) instead of nuke-and-rebuild** ([`366a9ad`](https://github.com/techempower-org/mempalace/commit/366a9ad))
   Fork main had been carrying ``cmd_purge``'s nuke-and-rebuild
   shape (extract survivors, ``shutil.rmtree``, recreate, re-insert).
   Cherry-picked the post-review rewrite from PR #1087's branch:
@@ -1506,7 +1506,7 @@ config-file readback. Suite total 1562 passed.
 ### Fixed
 
 
-- **Integrity gate prevents quarantine_stale_hnsw from destroying healthy indexes** ([`645ba20`](https://github.com/jphein/mempalace/commit/645ba20))
+- **Integrity gate prevents quarantine_stale_hnsw from destroying healthy indexes** ([`645ba20`](https://github.com/techempower-org/mempalace/commit/645ba20))
   Previous behavior fired whenever ``sqlite_mtime - hnsw_mtime`` exceeded
   the (lowered, in #1173) 300s threshold. ChromaDB 1.5.x flushes HNSW
   asynchronously and a clean shutdown does not force-flush, so the
@@ -1530,7 +1530,7 @@ config-file readback. Suite total 1562 passed.
 ### Performance
 
 
-- **Cherry-pick #1085 — batch ChromaDB inserts in miner (10–30× faster)** ([`6be6fff`](https://github.com/jphein/mempalace/commit/6be6fff))
+- **Cherry-pick #1085 — batch ChromaDB inserts in miner (10–30× faster)** ([`6be6fff`](https://github.com/techempower-org/mempalace/commit/6be6fff))
   Cherry-picked from upstream PR
   [#1085](https://github.com/MemPalace/mempalace/pull/1085) (@midweste,
   OPEN as of 2026-04-26). New ``_build_drawer()`` helper + ``add_drawers()``
@@ -1539,10 +1539,13 @@ config-file readback. Suite total 1562 passed.
   and ``os.path.getmtime()`` to file-level (2 syscalls per file instead
   of 2N). Reported 10–30× mining speedup upstream. Fork-side resolution
   preserved fork's existing ``DRAWER_UPSERT_BATCH_SIZE=1000``; aliased
-  upstream's ``CHROMA_BATCH_LIMIT`` to it. Becomes a no-op when #1085
-  merges to develop and we next sync.
+  upstream's ``CHROMA_BATCH_LIMIT`` to it. **2026-05-16:** #1085 was
+  closed by @midweste, superseded by merged upstream
+  [#1185](https://github.com/MemPalace/mempalace/pull/1185) (wider
+  scope: same batching + optional GPU acceleration). The fork-side
+  cherry-pick is now a no-op against develop; drop on next sync.
 
-  *Upstream:* [PR #1085](https://github.com/MemPalace/mempalace/pull/1085) (OPEN)
+  *Upstream:* [PR #1085](https://github.com/MemPalace/mempalace/pull/1085) (CLOSED)
   *Files:* `mempalace/miner.py`
 
 
@@ -1552,7 +1555,7 @@ config-file readback. Suite total 1562 passed.
 ### Added
 
 
-- **Phases A–C of the checkpoint collection split** ([`e266365`](https://github.com/jphein/mempalace/commit/e266365))
+- **Phases A–C of the checkpoint collection split** ([`e266365`](https://github.com/techempower-org/mempalace/commit/e266365))
   New ``mempalace_session_recovery`` collection adapter
   (``_SESSION_RECOVERY_COLLECTION`` + ``get_session_recovery_collection``
   in ``palace.py``); ``tool_diary_write`` routes ``topic in _CHECKPOINT_TOPICS``
@@ -1570,7 +1573,7 @@ config-file readback. Suite total 1562 passed.
 ### Fixed
 
 
-- **Gate quarantine_stale_hnsw to once-per-palace-per-process** ([`70c4bc6`](https://github.com/jphein/mempalace/commit/70c4bc6))
+- **Gate quarantine_stale_hnsw to once-per-palace-per-process** ([`70c4bc6`](https://github.com/techempower-org/mempalace/commit/70c4bc6))
   ``make_client()`` previously invoked ``quarantine_stale_hnsw`` on every
   reconnect; under steady write load the proactive check kept firing,
   racking up ``.drift-*`` directories every 10–30 minutes. New
@@ -1585,7 +1588,7 @@ config-file readback. Suite total 1562 passed.
   *Files:* `mempalace/backends/chroma.py`, `tests/test_backends.py`, `tests/conftest.py`
 
 
-- **palace_graph.build_graph skips None metadata** ([`5fd15db`](https://github.com/jphein/mempalace/commit/5fd15db))
+- **palace_graph.build_graph skips None metadata** ([`5fd15db`](https://github.com/techempower-org/mempalace/commit/5fd15db))
   ``palace_graph.py:95`` was calling ``meta.get("room", "")`` unconditionally;
   ChromaDB returns ``None`` for legacy/partial-write drawers, taking out
   every consumer of ``build_graph`` (graph_stats, find_tunnels, traverse,
@@ -1597,7 +1600,7 @@ config-file readback. Suite total 1562 passed.
   *Files:* `mempalace/palace_graph.py`
 
 
-- **kind= filter on search_memories excludes Stop-hook checkpoints (transitional)** ([`f9f5cc4`](https://github.com/jphein/mempalace/commit/f9f5cc4))
+- **kind= filter on search_memories excludes Stop-hook checkpoints (transitional)** ([`f9f5cc4`](https://github.com/techempower-org/mempalace/commit/f9f5cc4))
   Three values: ``"content"`` (default, excludes), ``"checkpoint"``
   (recovery/audit only), ``"all"`` (no filter). Two same-day architecture
   corrections: (a) the where-clause filter (``topic $nin [...]``) tripped
@@ -1625,6 +1628,7 @@ config-file readback. Suite total 1562 passed.
 *See CHANGELOG.md (upstream) for the full released history.*
 
 
+- [PR #1024](https://github.com/MemPalace/mempalace/pull/1024) — Configurable chunk_size / chunk_overlap / min_chunk_size — 2026-05-15
 - [PR #1173](https://github.com/MemPalace/mempalace/pull/1173) — quarantine_stale_hnsw on make_client + cold-start gate + integrity sniff — 2026-04-26
 - [PR #1177](https://github.com/MemPalace/mempalace/pull/1177) — `.blob_seq_ids_migrated` marker guard (closes #1090) — 2026-04-26
 - [PR #1198](https://github.com/MemPalace/mempalace/pull/1198) — _tokenize None-document guard in BM25 reranker — 2026-04-26
diff --git a/README.md b/README.md
index 8e7c70c910..dcff528f53 100644
--- a/README.md
+++ b/README.md
@@ -25,7 +25,7 @@
 
 ---
 
-This fork tracks `upstream/develop` through the 2026-05-10 sync (commit `a67be3f`, fork merge via `chore/sync-develop-2026-05-10`) and runs in production on a **273K-drawer Postgres + pgvector + Apache AGE palace** behind [palace-daemon](https://github.com/techempower-org/palace-daemon) on `disks:8085`. It carries ~434 fork-ahead changes that compose with — not replace — bensig's release direction; the v3.3.5 release (2026-05-10) includes our co-authored `_get_collection` retry-once via upstream #1377. ~1850 tests pass on `main`. The new things here are *what we've learned*, not just what we've fixed.
+This fork tracks `upstream/develop` through the 2026-05-10 sync (commit `a67be3f`, fork merge via `chore/sync-develop-2026-05-10`) and runs in production on a **300K+ drawer Postgres + pgvector + Apache AGE palace** behind [palace-daemon](https://github.com/techempower-org/palace-daemon) on `disks:8085`. It carries ~434 fork-ahead changes that compose with — not replace — bensig's release direction; the v3.3.5 release (2026-05-10) includes our co-authored `_get_collection` retry-once via upstream #1377. 2486 tests pass on `main`. The new things here are *what we've learned*, not just what we've fixed.
 
 **2026-05-15 substrate cutover complete.** What was on the `feat/pgvector-age-impl` branch is now on `main` and serving production traffic. **Hybrid retrieval** (`candidate_strategy="hybrid"` — vector ∪ postgres-`tsvector` BM25 ∪ AGE graph-expanded candidates, hybrid-reranked) shipped end-to-end through `palace-daemon`'s new `/search/keyword` and `/search/hybrid` endpoints. **8 of 9 bench suites pass** on the postgres backend (`test_chromadb_stress` skipped, doesn't apply) — no regressions on the search heap hot path vs the chromadb-era numbers. See [`docs/operators/pgvector-cutover-runbook.md`](docs/operators/pgvector-cutover-runbook.md) for the operator narrative and [`docs/benchmarks/2026-05-15-remaining-benches.json`](docs/benchmarks/2026-05-15-remaining-benches.json) for the bench artifact.
 
@@ -101,7 +101,7 @@ We surveyed the memory-system landscape in April 2026 and found no verbatim-firs
 
 | System | Verbatim? | Local? | MCP? | First public | Notes |
 |---|---|---|---|---|---|
-| **MemPalace** | Yes | Yes | Yes | 2026-04-06 (v3.0.0) | What we have. ~160K drawers in production. Verbatim drawers + wings/rooms scope + SQLite KG + BM25/vector hybrid search. |
+| **MemPalace** | Yes | Yes | Yes | 2026-04-06 (v3.0.0) | What we have. 300K+ drawers in production. Verbatim drawers + wings/rooms scope + SQLite KG + BM25/vector hybrid search. |
 | [Longhand](https://glama.ai/mcp/servers/Wynelson94/longhand) | Yes | Yes | Yes, 17-tool MCP | 2026-04-14 (v0.5.5) | Closest cousin. Claude Code-specific — reads `~/.claude/projects/*.jsonl` directly. SQLite (raw JSON per event) + ChromaDB (embeddings of pre-computed "episodes"). Deterministic file-state replay via stored diffs. |
 | [Celiums](https://celiums.ai/) | Yes | Yes (SQLite, Docker, or DO) | Yes, 6-tool MCP | 2026-04-08 (repo) | Stores full module text with PAD emotional vectors, importance scores, and circadian metadata. Bundles a 500K+ expert-module knowledge base alongside personal memory — different product shape. |
 | [mcp-memory-service](https://github.com/doobidoo/mcp-memory-service) | Yes by default (opt-in consolidation) | Yes (SQLite) or Cloudflare Workers | Yes | 2024-12-26 | The long-standing verbatim option. Turn-level storage; MiniLM embeddings local. Targets LangGraph / CrewAI / AutoGen plus Claude. |
@@ -116,7 +116,7 @@ The April-2026 verbatim cluster (MemPalace, Celiums, Longhand, engram all within
 
 ## Substrate: Postgres + pgvector + Apache AGE
 
-*Status (2026-05-15): **shipped on `main`** and serving production traffic. Postgres backend cherry-picked from upstream [#665](https://github.com/MemPalace/mempalace/pull/665) (skuznetsov's PostgreSQL backend on the RFC 001 `BaseBackend` contract) plus our pgvector lazy-index race fix (commit `4566f8a`, also posted to #665 as an operator follow-up). Live on `disks.jphe.in:5433` running PG16 + pgvector 0.8.2 + AGE 1.6.0 via `apache/age:release_PG16_1.6.0` + `postgresql-16-pgvector`. **273K-drawer palace.** Hybrid retrieval (vector ∪ tsvector BM25 ∪ AGE graph) shipped end-to-end as `candidate_strategy="hybrid"`, exposed via `palace-daemon`'s `/search/keyword` and `/search/hybrid` endpoints. 8/9 mempalace bench suites pass on postgres backend ([`docs/benchmarks/2026-05-15-remaining-benches.json`](docs/benchmarks/2026-05-15-remaining-benches.json)). Original WAIT-for-#665-upstream-merge stance superseded; the cutover happened on our timeline. Full operator narrative at [`docs/operators/pgvector-cutover-runbook.md`](docs/operators/pgvector-cutover-runbook.md).*
+*Status (2026-05-15): **shipped on `main`** and serving production traffic. Postgres backend cherry-picked from upstream [#665](https://github.com/MemPalace/mempalace/pull/665) (skuznetsov's PostgreSQL backend on the RFC 001 `BaseBackend` contract) plus our pgvector lazy-index race fix (commit `4566f8a`, also posted to #665 as an operator follow-up). Live on `disks.jphe.in:5433` running PG16 + pgvector 0.8.2 + AGE 1.6.0 via `apache/age:release_PG16_1.6.0` + `postgresql-16-pgvector`. **300K+ drawer palace.** Hybrid retrieval (vector ∪ tsvector BM25 ∪ AGE graph) shipped end-to-end as `candidate_strategy="hybrid"`, exposed via `palace-daemon`'s `/search/keyword` and `/search/hybrid` endpoints. 8/9 mempalace bench suites pass on postgres backend ([`docs/benchmarks/2026-05-15-remaining-benches.json`](docs/benchmarks/2026-05-15-remaining-benches.json)). Original WAIT-for-#665-upstream-merge stance superseded; the cutover happened on our timeline. Full operator narrative at [`docs/operators/pgvector-cutover-runbook.md`](docs/operators/pgvector-cutover-runbook.md).*
 
 This is composition, not a fork-led architectural shift: `BaseBackend` + `BaseCollection` + `PalaceRef` + the entry-point registry already live in upstream develop at [`mempalace/backends/`](https://github.com/MemPalace/mempalace/tree/develop/mempalace/backends), explicitly designed so third-party backends register via Python entry points without touching core. The architectural decision was upstream's; the fork's contribution is choosing pgvector + AGE as one specific implementation worth picking — and running it.
 
@@ -141,7 +141,7 @@ A six-phase plan to fully integrate the mempalace knowledge base with Apache AGE
 | 1 | `KnowledgeGraphAGE` API parity with the SQLite KG (6 new methods) | `mempalace/knowledge_graph_age.py` |
 | 2 | Write-through middleware on `PostgresCollection.add/upsert` (every drawer write extracts entities + creates `:MENTIONS` edges in AGE) | `mempalace/kg_writethrough.py` + `mempalace/backends/postgres.py` |
 | 3 | Palace structure as native AGE nodes (Wing → Room → Drawer + tunnels) | `mempalace/palace_graph_age.py` |
-| 4 | `backfill_age` — restartable, checkpointed one-shot for existing drawer tables (~22h for production 274K palace) | `mempalace/backfill_age.py` |
+| 4 | `backfill_age` — restartable, checkpointed one-shot for existing drawer tables (~22h for production 300K+ palace) | `mempalace/backfill_age.py` |
 | 5 | `POST /search/age-fused` endpoint on palace-daemon — vector ⊕ AGE graph RRF fusion | (palace-daemon repo) |
 | 6 | `mempalace_walk_palace` MCP tool — `start_wing` / `start_room` / `start_entity` × depth | `mempalace/mcp_server.py` |
 
@@ -168,7 +168,7 @@ Independent same-day cross-fork verification by [@nakata-app](https://github.com
 Three bands of work, all instances of the principles above. Detail rows in the [appendix](#fork-change-inventory) at the bottom.
 
 - **Structural retrieval fixes (Principle 1, Principle 2).** Verbatim-only model: hooks no longer write 1KB checkpoint summaries; auto-mined transcript chunks land in `mempalace_drawers` and `mempalace_search` reaches them directly. The earlier dedicated `mempalace_session_recovery` collection (Apr 25–May 5) and its read-only `mempalace_session_recovery_read` MCP tool have been retired (May 5 — see `docs/superpowers/specs/2026-05-05-verbatim-only-design.md`). Net result: one collection, one search path, no kind-filter / over-fetch hack. `drawer_id` surfacing on every search/diary hit so callers can build citation popovers and follow-ups.
-- **Single-writer architecture (Principle 3).** [palace-daemon](https://github.com/jphein/palace-daemon) is the only process that opens the palace; clients connect over HTTP. ChromaDB 1.5.x's HNSW concurrency hazards (`#974`/`#965`/`#823` family) become structurally impossible. Cold-start integrity sniff-test on segment metadata files prevents `quarantine_stale_hnsw` from destroying healthy indexes during async-flush lag. Cherry-pick of upstream [#1085](https://github.com/MemPalace/mempalace/pull/1085) for 10–30× mining speedup; fork-side boundary-level None-metadata coercion that closes a per-site-guard family, filed upstream as [#1094](https://github.com/MemPalace/mempalace/pull/1094).
+- **Single-writer architecture (Principle 3).** [palace-daemon](https://github.com/techempower-org/palace-daemon) is the only process that opens the palace; clients connect over HTTP. ChromaDB 1.5.x's HNSW concurrency hazards (`#974`/`#965`/`#823` family) become structurally impossible. Cold-start integrity sniff-test on segment metadata files prevents `quarantine_stale_hnsw` from destroying healthy indexes during async-flush lag. Cherry-pick of upstream [#1085](https://github.com/MemPalace/mempalace/pull/1085) for 10–30× mining speedup; fork-side boundary-level None-metadata coercion that closes a per-site-guard family, filed upstream as [#1094](https://github.com/MemPalace/mempalace/pull/1094).
 - **Deterministic hook saves (Principles 1+2+3 compose).** Silent saves bypass auto-memory conflicts entirely — the LLM is no longer in the save path, so `decision: "block"` race conditions and Claude's auto-memory winning over MCP tools both go away. Verbatim transcript ingest is the entire save path; the save marker advances on each fire and `systemMessage` reports the wing the ingest landed in. PreCompact does the same — sync-mines the transcript before context boundary, no separate marker write.
 
 ## Quickstart
@@ -183,7 +183,7 @@ uv run mempalace mine ~/Projects/myproject
 uv run mempalace search "why did we switch to GraphQL"
 ```
 
-For a daemon-fronted deployment (recommended once palace size reaches the multi-thousand-drawer range), see [palace-daemon](https://github.com/jphein/palace-daemon)'s setup. The fork's `scripts/deploy.sh` is a one-command Syncthing-aware redeploy: push fork main, restart palace-daemon, post-restart import-check that the new fork-ahead surface is loaded.
+For a daemon-fronted deployment (recommended once palace size reaches the multi-thousand-drawer range), see [palace-daemon](https://github.com/techempower-org/palace-daemon)'s setup. The fork's `scripts/deploy.sh` is a one-command Syncthing-aware redeploy: push fork main, restart palace-daemon, post-restart import-check that the new fork-ahead surface is loaded.
 
 ## What it looks like in production
 
@@ -229,7 +229,7 @@ Same instinct as the verbatim-vs-derivative axis. Derivative work belongs *next
 Hierarchy works when it's derived from unambiguous signals (cwd, transcript path, project directory) — not when it's hand-classified by content inspection. The earlier mistake was conflating "hierarchy is bad" with "mandatory synchronous classification is bad" — different claims.
 
 **Good uses of hierarchy, which we keep:**
-- **Browseable scope** for serendipitous recall across ~160K drawers.
+- **Browseable scope** for serendipitous recall across 300K+ drawers.
 - **Deletion and retention as a unit.** Purging an abandoned project is one operation, not a risky query-then-delete.
 - **Disambiguation without query gymnastics.** The same keyword across years of unrelated work.
 - **Auto-surfacing priors.** A wing derived from cwd is a cheap, unambiguous scoping signal.
@@ -253,7 +253,7 @@ Claude Code has two complementary memory layers, used in tandem:
 | Layer | Storage | Size | Consolidation | Purpose |
 |---|---|---|---|---|
 | **Auto-memory** | `~/.claude/projects/*/memory/*.md` | ~17 files (this project) | **Auto Dream** (research preview in both Claude Code and the Managed Agents Dreams API; beta header `dreaming-2026-04-21`) | Preferences, feedback, context |
-| **MemPalace** | palace-daemon at `http://disks.jphe.in:8085` (ChromaDB on the daemon host) | ~160K drawers | None — deliberately. See below. | Verbatim conversations, tool output, code |
+| **MemPalace** | palace-daemon at `http://disks.jphe.in:8085` (postgres + pgvector + Apache AGE) | 300K+ drawers | None — deliberately. See below. | Verbatim conversations, tool output, code |
 
 **Auto Dream's arrival ratifies the verbatim-vs-derivative axis.** Anthropic shipped Auto Dream in two surfaces in late April, both research preview: a [research-preview consolidator inside Claude Code](https://claudefa.st/blog/guide/mechanics/auto-dream) (manual `/dream` or auto-trigger when both 24 hours and 5+ sessions have elapsed since the last run; reads `~/.claude/projects/<project>/memory/` + MEMORY.md + JSONL transcripts + daily logs; writes back to the same memory directory) and a [Dreams API in Managed Agents](https://platform.claude.com/docs/en/managed-agents/dreams) (REST research preview, beta header `dreaming-2026-04-21`, supported models `claude-opus-4-7` and `claude-sonnet-4-6`, up to 100 sessions per dream, optional 4096-character `instructions`).
 
@@ -316,7 +316,7 @@ Knowledge lives across 7+ layers: global CLAUDE.md, project CLAUDE.md, auto-memo
 
 A meaningful shift in 2026-04 and 2026-05: this fork increasingly *composes with* upstream rather than carrying parallel implementations.
 
-- **Cherry-picks (in-flight upstream PRs we use early):** [#665](https://github.com/MemPalace/mempalace/pull/665) PostgreSQL backend (commit `5e90c72`, the substrate work above), [#1085](https://github.com/MemPalace/mempalace/pull/1085) batched inserts (`6be6fff` — CLOSED 2026-05-05, superseded by merged [#1185](https://github.com/MemPalace/mempalace/pull/1185); safe to drop on next sync), [#1087 rewrite](https://github.com/MemPalace/mempalace/pull/1087) `cmd_purge` via `delete(where=)` (`366a9ad`), [#1094](https://github.com/MemPalace/mempalace/pull/1094) None-metadata coercion (`43d728d`).
+- **Cherry-picks (in-flight upstream PRs we use early):** [#665](https://github.com/MemPalace/mempalace/pull/665) PostgreSQL backend (commit `5e90c72`, the substrate work above), [#1085](https://github.com/MemPalace/mempalace/pull/1085) batched inserts (`6be6fff` — CLOSED 2026-05-16, superseded by merged [#1185](https://github.com/MemPalace/mempalace/pull/1185); safe to drop on next sync), [#1087 rewrite](https://github.com/MemPalace/mempalace/pull/1087) `cmd_purge` via `delete(where=)` (`366a9ad`), [#1094](https://github.com/MemPalace/mempalace/pull/1094) None-metadata coercion (`43d728d`).
 - **Co-authored merges:** [#1377](https://github.com/MemPalace/mempalace/pull/1377) (surgical `_get_collection` retry-once, shipped in v3.3.5 — originated from this fork via #1286 which igorls closed and re-extracted with `Co-authored-by` credit).
 - **Coordinated reviews:** [#1199](https://github.com/MemPalace/mempalace/pull/1199) (rmdes' unbounded-ingest fix — pulled and tested locally, +1 with composition note), [#1219](https://github.com/MemPalace/mempalace/pull/1219) (pepo72's drawer_id — narrower than ours; offered the diary/recovery extension), [RFC 001 #743](https://github.com/MemPalace/mempalace/pull/743) (storage backend spec — flagged the multi-collection-by-purpose pattern as worth naming explicitly).
 - **Closed in favor of upstream:** [#1171](https://github.com/MemPalace/mempalace/pull/1171) cross-process write lock (closed 2026-04-25 — Felipe's [#976](https://github.com/MemPalace/mempalace/pull/976) `mine_global_lock` at the right layer plus daemon-strict architecture obsoleted ours).
@@ -329,7 +329,7 @@ From a 2026-04-21 sweep of upstream MemPalace issue + comment + discussion histo
 
 ### Companion tools (compose with MemPalace, don't replace it)
 
-- **[palace-daemon](https://github.com/rboarescu/palace-daemon)** (@rboarescu) — FastAPI gateway + MCP-over-HTTP proxy. Three asyncio semaphores (read / write / mine). Pins correctness floor at MemPalace ≥3.3.2. **This fork migrated to palace-daemon on 2026-04-24** ([`c09582c`](https://github.com/jphein/mempalace/commit/c09582c) wired MCP + hooks; [`0e97b19`](https://github.com/jphein/mempalace/commit/0e97b19) added daemon-strict mode). All reads and writes from the plugin flow through the daemon; auto-migrate-on-startup of the checkpoint split landed as palace-daemon [`034023c`](https://github.com/jphein/palace-daemon/commit/034023c). JP's deployment runs at [`jphein/palace-daemon`](https://github.com/jphein/palace-daemon).
+- **[palace-daemon](https://github.com/rboarescu/palace-daemon)** (@rboarescu) — FastAPI gateway + MCP-over-HTTP proxy. Three asyncio semaphores (read / write / mine). Pins correctness floor at MemPalace ≥3.3.2. **This fork migrated to palace-daemon on 2026-04-24** ([`c09582c`](https://github.com/techempower-org/mempalace/commit/c09582c) wired MCP + hooks; [`0e97b19`](https://github.com/techempower-org/mempalace/commit/0e97b19) added daemon-strict mode). All reads and writes from the plugin flow through the daemon; auto-migrate-on-startup of the checkpoint split landed as palace-daemon [`034023c`](https://github.com/techempower-org/palace-daemon/commit/034023c). JP's deployment runs at [`techempower-org/palace-daemon`](https://github.com/techempower-org/palace-daemon).
 - **[engram](https://github.com/NickCirv/engram)** (@NickCirv) — File-read interception for AI coding assistants. Uses MemPalace as one of six context providers via `mcp-mempalace mempalace-search`; caches with 1h TTL. Upstream [discussion #798](https://github.com/MemPalace/mempalace/discussions/798).
 - **[engram](https://github.com/harreh3iesh/engram)** (@harreh3iesh — different project, same name) — Hooks + tools for AI memory, first-class MemPalace backend. **Stuck detector** (`PreToolUse` hook counts Grep/Glob calls and nudges the AI when spinning) is a pattern worth borrowing. Upstream [discussion #748](https://github.com/MemPalace/mempalace/discussions/748).
 - **[cdd-mempalace](https://github.com/fuzzymoomoo/cdd-mempalace)** (@fuzzymoomoo) — Bridge library mapping Context-Driven Development methodology onto wings/halls/rooms. Multiple active upstream PRs.
@@ -358,7 +358,7 @@ Built *on top of* or *alongside* MemPalace by community contributors who use the
 - **[Kent](https://github.com/kenchambers/kent)** (@kenchambers) — typed async agent runtime with APO training subsystem and recall games.
 - **[adaptmem](https://github.com/nakata-app/adaptmem)** (@nakata-app) — encoder fine-tune adapter; orthogonal lift across retrieval modes.
 - **[GraphPalace](https://github.com/web3guru888/GraphPalace)** (@web3guru888) — graph-layer build. Forked at [jphein/GraphPalace](https://github.com/jphein/GraphPalace).
-- **[mempalace-viz](https://github.com/JoeDoesJits/mempalace-viz)** (@JoeDoesJits) — visualization layer (wings, rooms, tunnels, drawer counts). Forked at [jphein/mempalace-viz](https://github.com/jphein/mempalace-viz).
+- **[mempalace-viz](https://github.com/JoeDoesJits/mempalace-viz)** (@JoeDoesJits) — visualization layer (wings, rooms, tunnels, drawer counts). Forked at [jphein/mempalace-viz](https://github.com/techempower-org/mempalace-viz).
 - **[AutomataArena](https://github.com/astrutt/AutomataArena)** (@astrutt) — multi-agent orchestration substrate. Forked at [jphein/AutomataArena](https://github.com/jphein/AutomataArena).
 
 ### Active forks beyond ours
@@ -373,13 +373,12 @@ Built *on top of* or *alongside* MemPalace by community contributors who use the
 
 ## Open upstream PRs
 
-Ten open from this fork as of 2026-05-12. Run `gh pr list --repo MemPalace/mempalace --author jphein --state open` for the live list. Two merged today: [#1459](https://github.com/MemPalace/mempalace/pull/1459) (empty-metadata sentinel) and [#1474](https://github.com/MemPalace/mempalace/pull/1474) (convo_miner bulk pre-fetch).
+Open from this fork as of 2026-05-22. Run `gh pr list --repo MemPalace/mempalace --author jphein --state open` for the live list. Recently merged: [#1024](https://github.com/MemPalace/mempalace/pull/1024) (configurable chunking, 2026-05-15), [#1459](https://github.com/MemPalace/mempalace/pull/1459) (empty-metadata sentinel, 2026-05-12) and [#1474](https://github.com/MemPalace/mempalace/pull/1474) (convo_miner bulk pre-fetch, 2026-05-12).
 
 | PR | Status | Description |
 |---|---|---|
 | [#660](https://github.com/MemPalace/mempalace/pull/660) | CI green, awaiting review | L1 importance pre-filter |
 | [#1005](https://github.com/MemPalace/mempalace/pull/1005) | CI green, Dialectician-acked | Warnings + sqlite BM25 top-up — never silently return fewer results than scope contains |
-| [#1024](https://github.com/MemPalace/mempalace/pull/1024) | CI green, qodo-acked | Configurable `chunk_size` / `chunk_overlap` / `min_chunk_size` |
 | [#1086](https://github.com/MemPalace/mempalace/pull/1086) | CI green, awaiting review | `mempalace export` CLI wrapper |
 | [#1087](https://github.com/MemPalace/mempalace/pull/1087) | CI green, **rewritten 2026-04-26** per @igorls's review | `mempalace purge --wing/--room` via `delete(where=)` (no nuke-and-rebuild) |
 | [#1094](https://github.com/MemPalace/mempalace/pull/1094) | CI green, awaiting review | Coerce `None` metadatas to `{}` at `ChromaCollection` boundary |
@@ -432,17 +431,15 @@ The canonical source is [`docs/fork-changes.yaml`](docs/fork-changes.yaml); [`FO
 
 | Area | Change | Status | Files |
 |---|---|---|---|
-| **Backend** | **Postgres + pgvector + Apache AGE substrate implementation.** Cherry-pick of upstream [#665](https://github.com/MemPalace/mempalace/pull/665) (skuznetsov's PostgreSQL backend on the RFC 001 `BaseBackend` contract). Smoke tests + BaseCollection conformance test added. Live test container on JP's homelab LAN (`10.0.6.120:5433`) running PG16 + pgvector 0.8.2 + AGE 1.6.0. Composition stance: WAIT for #665 to merge; Plan-B trigger 2026-06-08. | In-flight on `feat/pgvector-age-impl` — PR [#21](https://github.com/jphein/mempalace/pull/21). Fork commits `5e90c72` (cherry-pick), `04c6294` (smoke), `1d5486f` (docstring cleanup). | `mempalace/backends/postgres.py`, `mempalace/palace.py`, `tests/test_backends_postgres.py`, `docs/internal/pgvector-665-decision.md` |
-| **Reliability** | **Daemon-strict migration completion** (May 7). Closes the last desktop-side write paths that bypassed palace-daemon. `mcp_server.py` gates at the `handle_request` JSON-RPC chokepoint and forwards every method to the daemon's `/mcp` proxy when `PALACE_DAEMON_URL` is set; `cli.py` gates `cmd_status`, `cmd_search`, `cmd_mine` against the same env var. Mirrors the gate `hooks_cli.py` already uses (2026-04-24, drift-incident fix). With the local `mempalace-data/` no longer pinned by `~/.mempalace/config.json`, the canonical palace at `disks.jphe.in:8085` is the single writer for every desktop entry point. CLI `init`/`repair`/`export`/`sweep`/`purge`/`mined`/`wakeup` stay local because they need on-host filesystem access. | Fork-ahead, pitchable upstream as a single-file replacement for `palace-daemon/clients/mempalace-mcp.py`. Fork commits [`41359ba`](https://github.com/jphein/mempalace/commit/41359ba) (mcp_server) + [`22ef562`](https://github.com/jphein/mempalace/commit/22ef562) (CLI). | `mcp_server.py`, `cli.py`, `tests/conftest.py`, `tests/test_mcp_server_daemon.py`, `tests/test_cli_daemon.py` |
-| **Search** | **Verbatim-only retrieval** (May 5). Hooks write only verbatim transcript chunks; the dedicated `mempalace_session_recovery` collection and `mempalace_session_recovery_read` MCP tool are retired. `mempalace_search` reaches all session content directly. Replaces the earlier multi-collection split (Apr 25 → May 5). Spec: `docs/superpowers/specs/2026-05-05-verbatim-only-design.md`. | PRs in review — [`#2`](https://github.com/jphein/mempalace/pull/2) (transcript ingest restore), [`#3`](https://github.com/jphein/mempalace/pull/3) (drop checkpoint writes), [`#5`](https://github.com/jphein/mempalace/pull/5) (retire collection); palace-daemon [`#1`](https://github.com/jphein/palace-daemon/pull/1) (path translation) | `hooks_cli.py`, `mcp_server.py`, `palace.py`, `migrate.py`, `cli.py` |
-| **Search** | Surface `drawer_id` in `mempalace_search` results and `mempalace_diary_read` entries. ChromaDB primary key was returned but never plumbed into the result-building loop. | PR pending — fork commit [`9a8bb77`](https://github.com/jphein/mempalace/commit/9a8bb77); upstream [#1219](https://github.com/MemPalace/mempalace/pull/1219) (@pepo72) is the narrower searcher-only equivalent. | `searcher.py`, `mcp_server.py`, `tests/...`, `website/reference/mcp-tools.md` |
-| **CLI** | `mempalace mined` lists mined source files grouped by wing × source_file; `mempalace purge --source-file` deletes drawers from a specific file. Closes the "removing manually mined data" half of the mining-management ask. | [`#4`](https://github.com/jphein/mempalace/pull/4) | `cli.py`, `tests/test_cli.py` |
-| **Performance** | Cherry-picked upstream [#1085](https://github.com/MemPalace/mempalace/pull/1085) (@midweste) — batch ChromaDB inserts in miner. New `_build_drawer()` + `add_drawers()`. Reported 10–30× mining speedup. | Cherry-pick — fork commit [`6be6fff`](https://github.com/jphein/mempalace/commit/6be6fff). **2026-05-16:** #1085 closed by author, superseded by merged upstream [#1185](https://github.com/MemPalace/mempalace/pull/1185) (wider scope: same batching + optional GPU acceleration). Our cherry-pick is now a no-op against develop; drop on next sync. | `mempalace/miner.py` |
-| **Reliability** | Coerce None metadatas at chromadb boundary. Closes the per-site-guard family of None-metadata bugs (#999, #1198, #1201) at one site instead of N. Fork-authored; filed upstream as [#1094](https://github.com/MemPalace/mempalace/pull/1094). | Fork commit [`43d728d`](https://github.com/jphein/mempalace/commit/43d728d); upstream PR open | `backends/chroma.py`, `tests/test_backends.py` |
+| **Backend** | **Postgres + pgvector + Apache AGE substrate implementation.** Cherry-pick of upstream [#665](https://github.com/MemPalace/mempalace/pull/665) (skuznetsov's PostgreSQL backend on the RFC 001 `BaseBackend` contract). Smoke tests + BaseCollection conformance test added. Live test container on JP's homelab LAN (`10.0.6.120:5433`) running PG16 + pgvector 0.8.2 + AGE 1.6.0. Composition stance: WAIT for #665 to merge; Plan-B trigger 2026-06-08. | In-flight on `feat/pgvector-age-impl` — PR [#21](https://github.com/techempower-org/mempalace/pull/21). Fork commits `5e90c72` (cherry-pick), `04c6294` (smoke), `1d5486f` (docstring cleanup). | `mempalace/backends/postgres.py`, `mempalace/palace.py`, `tests/test_backends_postgres.py`, `docs/internal/pgvector-665-decision.md` |
+| **Reliability** | **Daemon-strict migration completion** (May 7). Closes the last desktop-side write paths that bypassed palace-daemon. `mcp_server.py` gates at the `handle_request` JSON-RPC chokepoint and forwards every method to the daemon's `/mcp` proxy when `PALACE_DAEMON_URL` is set; `cli.py` gates `cmd_status`, `cmd_search`, `cmd_mine` against the same env var. Mirrors the gate `hooks_cli.py` already uses (2026-04-24, drift-incident fix). With the local `mempalace-data/` no longer pinned by `~/.mempalace/config.json`, the canonical palace at `disks.jphe.in:8085` is the single writer for every desktop entry point. CLI `init`/`repair`/`export`/`sweep`/`purge`/`mined`/`wakeup` stay local because they need on-host filesystem access. | Fork-ahead, pitchable upstream as a single-file replacement for `palace-daemon/clients/mempalace-mcp.py`. Fork commits [`41359ba`](https://github.com/techempower-org/mempalace/commit/41359ba) (mcp_server) + [`22ef562`](https://github.com/techempower-org/mempalace/commit/22ef562) (CLI). | `mcp_server.py`, `cli.py`, `tests/conftest.py`, `tests/test_mcp_server_daemon.py`, `tests/test_cli_daemon.py` |
+| **Search** | **Verbatim-only retrieval** (May 5). Hooks write only verbatim transcript chunks; the dedicated `mempalace_session_recovery` collection and `mempalace_session_recovery_read` MCP tool are retired. `mempalace_search` reaches all session content directly. Replaces the earlier multi-collection split (Apr 25 → May 5). Spec: `docs/superpowers/specs/2026-05-05-verbatim-only-design.md`. | PRs in review — [`#2`](https://github.com/techempower-org/mempalace/pull/2) (transcript ingest restore), [`#3`](https://github.com/techempower-org/mempalace/pull/3) (drop checkpoint writes), [`#5`](https://github.com/techempower-org/mempalace/pull/5) (retire collection); palace-daemon [`#1`](https://github.com/techempower-org/palace-daemon/pull/1) (path translation) | `hooks_cli.py`, `mcp_server.py`, `palace.py`, `migrate.py`, `cli.py` |
+| **Search** | Surface `drawer_id` in `mempalace_search` results and `mempalace_diary_read` entries. ChromaDB primary key was returned but never plumbed into the result-building loop. | PR pending — fork commit [`9a8bb77`](https://github.com/techempower-org/mempalace/commit/9a8bb77); upstream [#1219](https://github.com/MemPalace/mempalace/pull/1219) (@pepo72) is the narrower searcher-only equivalent. | `searcher.py`, `mcp_server.py`, `tests/...`, `website/reference/mcp-tools.md` |
+| **CLI** | `mempalace mined` lists mined source files grouped by wing × source_file; `mempalace purge --source-file` deletes drawers from a specific file. Closes the "removing manually mined data" half of the mining-management ask. | [`#4`](https://github.com/techempower-org/mempalace/pull/4) | `cli.py`, `tests/test_cli.py` |
+| **Reliability** | Coerce None metadatas at chromadb boundary. Closes the per-site-guard family of None-metadata bugs (#999, #1198, #1201) at one site instead of N. Fork-authored; filed upstream as [#1094](https://github.com/MemPalace/mempalace/pull/1094). | Fork commit [`43d728d`](https://github.com/techempower-org/mempalace/commit/43d728d); upstream PR open | `backends/chroma.py`, `tests/test_backends.py` |
 | **CLI** | `mempalace purge --wing/--room` via `collection.delete(where=...)`. Earlier nuke-and-rebuild draft predicated on #521's race; @igorls's review traced the stack. Simpler version preserves embedding fn, no rmtree window, routes through `ChromaBackend`. | [#1087](https://github.com/MemPalace/mempalace/pull/1087), rewritten 2026-04-26 per review | `cli.py`, `tests/test_cli.py` |
 | **CLI** | `mempalace export` CLI wrapper for upstream's existing `export_palace()`. | [#1086](https://github.com/MemPalace/mempalace/pull/1086) | `cli.py` |
 | **Performance** | L1 importance pre-filter — `importance >= 3` first, full scan fallback. | [#660](https://github.com/MemPalace/mempalace/pull/660) | `layers.py` |
-| **Config** | Configurable chunking parameters — `chunk_size` (800), `chunk_overlap` (100), `min_chunk_size` (50) in `config.json`, exposed via `MempalaceConfig`. | [#1024](https://github.com/MemPalace/mempalace/pull/1024) | `config.py`, `miner.py`, `convo_miner.py` |
 | **Search** | Warnings + sqlite BM25 top-up when vector underdelivers — `search_memories` returns `warnings: [...]` + `available_in_scope`; fallback hits tagged `matched_via: "sqlite_bm25_fallback"`. The palace never silently returns fewer results than the scope contains. | [#1005](https://github.com/MemPalace/mempalace/pull/1005) | `searcher.py` |
 | **Docs** | `docs/RELEASING.md` with `mempalace-mcp` pre-release grep. | [#1142](https://github.com/MemPalace/mempalace/pull/1142), accepted by @bensig 2026-04-23 | `docs/RELEASING.md` |
 | **Hooks** | `mempal_save_hook.sh` Python auto-detection (`MEMPAL_PYTHON` → repo venv → system `python3`). Same pattern in `.claude-plugin/`. Replied on [#1049](https://github.com/MemPalace/mempalace/issues/1049) offering autodetect, awaiting maintainer arbitration on [#1069](https://github.com/MemPalace/mempalace/issues/1069). | PR pending after #1069 direction | `hooks/mempal_save_hook.sh`, `.claude-plugin/hooks/...` |
@@ -450,6 +447,7 @@ The canonical source is [`docs/fork-changes.yaml`](docs/fork-changes.yaml); [`FO
 
 ### Recently merged into upstream
 
+- **2026-05-15:** [#1024](https://github.com/MemPalace/mempalace/pull/1024) — Configurable `chunk_size` / `chunk_overlap` / `min_chunk_size` exposed via `MempalaceConfig`
 - **2026-05-06 (in v3.3.5):** [#1377](https://github.com/MemPalace/mempalace/pull/1377) — `_get_collection` retry-once + log-on-failure (~25 LOC + 2 tests; co-authored from this fork via the closed #1286)
 - **2026-05-01 (post-v3.3.4):** [#1262](https://github.com/MemPalace/mempalace/pull/1262) (get-then-create at chromadb backend boundary), [#1289](https://github.com/MemPalace/mempalace/pull/1289) (MCP-server-side companion), [#1303](https://github.com/MemPalace/mempalace/pull/1303) (`embedding_function=` plumbing — prevents ONNX SIGSEGV on Py3.14)
 - **2026-04-26:** [#1173](https://github.com/MemPalace/mempalace/pull/1173) (`quarantine_stale_hnsw` cold-start gate + integrity sniff), [#1177](https://github.com/MemPalace/mempalace/pull/1177) (`.blob_seq_ids_migrated` marker), [#1198](https://github.com/MemPalace/mempalace/pull/1198) (`_tokenize` None guard), [#1201](https://github.com/MemPalace/mempalace/pull/1201) (`palace_graph` None metadata)
@@ -461,6 +459,7 @@ The canonical source is [`docs/fork-changes.yaml`](docs/fork-changes.yaml); [`FO
 
 ### Closed (superseded or withdrawn)
 
+- [#1085](https://github.com/MemPalace/mempalace/pull/1085) (cherry-pick — closed by @midweste 2026-05-16, superseded by merged upstream [#1185](https://github.com/MemPalace/mempalace/pull/1185) with wider GPU-acceleration scope; fork's `6be6fff` is a no-op against develop, safe to drop on next sync)
 - [#1286](https://github.com/MemPalace/mempalace/pull/1286) (drifted; @igorls closed and re-extracted the surgical fix as #1377 with co-author credit)
 - [#1171](https://github.com/MemPalace/mempalace/pull/1171) (cross-process write lock — superseded by #976 + daemon-strict)
 - [#1146](https://github.com/MemPalace/mempalace/pull/1146) (duplicate of @igorls's [#1147](https://github.com/MemPalace/mempalace/pull/1147))
@@ -492,7 +491,7 @@ Articles, surveys, and research that shaped the fork's direction.
 - [**Claude Code Auto Dream guide**](https://claudefa.st/blog/guide/mechanics/auto-dream) — research-preview consolidation inside Claude Code; reads memory + JSONL, writes consolidated index.
 - [**Microsoft Tech Community — Combining pgvector and Apache AGE: knowledge graph & semantic intelligence in a single engine**](https://techcommunity.microsoft.com/blog/adforpostgresql/combining-pgvector-and-apache-age---knowledge-graph--semantic-intelligence-in-a-/4508781) (Raunak, 2026-04-15) — bridge-pattern reference for the substrate work.
 - [**Dave's Garage — "My Custom AI Went Superhuman Yesterday..."**](https://www.youtube.com/watch?v=TdbpoDjIvPk) (Dave Plummer, 2026-02-28) — conceptual reference for why graph structure matters in retrieval: vectors get you "topically nearby"; the graph gets you "actually related."
-- [**Phil Karlton's two hard things**](https://martinfowler.com/bliki/TwoHardThings.html) — naming and cache invalidation. Cited in "What this fork has learned" because, even at ~160K drawers, the day-to-day operational work is still mostly these two.
+- [**Phil Karlton's two hard things**](https://martinfowler.com/bliki/TwoHardThings.html) — naming and cache invalidation. Cited in "What this fork has learned" because, even at 300K+ drawers, the day-to-day operational work is still mostly these two.
 - [**Recursive Language Models**](https://arxiv.org/abs/2512.24601) (Zhang, Kraska, Khattab, 2025) — the RLM paper. The four-layer section's invocation-ceiling finding is measured against this paper's mechanism.
 - [**Think, But Don't Overthink: Reproducing Recursive Language Models**](https://arxiv.org/abs/2603.02615) — depth-2 collapse finding behind the recommendation to keep RLM at composition rather than retrieval.
 
diff --git a/docs/fork-changes.yaml b/docs/fork-changes.yaml
index 93466b8b31..d1b7ace002 100644
--- a/docs/fork-changes.yaml
+++ b/docs/fork-changes.yaml
@@ -1497,7 +1497,8 @@ entries:
       merges. Amends fork-ahead row 17.
     tests: "1548/1548 (was 1546/1548 with 2 isolation failures in test_convo_miner)"
     pr: 1024
-    pr_state: OPEN
+    pr_state: MERGED
+    merged_date: 2026-05-15
     files:
       - mempalace/config.py
     supersedes: []
@@ -1763,12 +1764,16 @@ entries:
       and ``os.path.getmtime()`` to file-level (2 syscalls per file instead
       of 2N). Reported 10–30× mining speedup upstream. Fork-side resolution
       preserved fork's existing ``DRAWER_UPSERT_BATCH_SIZE=1000``; aliased
-      upstream's ``CHROMA_BATCH_LIMIT`` to it. Becomes a no-op when #1085
-      merges to develop and we next sync.
+      upstream's ``CHROMA_BATCH_LIMIT`` to it. **2026-05-16:** #1085 was
+      closed by @midweste, superseded by merged upstream
+      [#1185](https://github.com/MemPalace/mempalace/pull/1185) (wider
+      scope: same batching + optional GPU acceleration). The fork-side
+      cherry-pick is now a no-op against develop; drop on next sync.
     pr: 1085
-    pr_state: OPEN
+    pr_state: CLOSED
     files:
       - mempalace/miner.py
+    supersedes: []
 
   - id: deploy-script
     date: 2026-04-26
@@ -1885,6 +1890,7 @@ merged_upstream:
     - "Trim entries from this list once they're more than ~30 days old."
     - "See CHANGELOG.md (upstream) for the full released history."
   entries:
+    - { pr: 1024, title: "Configurable chunk_size / chunk_overlap / min_chunk_size", merged: 2026-05-15 }
     - { pr: 1173, title: "quarantine_stale_hnsw on make_client + cold-start gate + integrity sniff", merged: 2026-04-26 }
     - { pr: 1177, title: "`.blob_seq_ids_migrated` marker guard (closes #1090)", merged: 2026-04-26 }
     - { pr: 1198, title: "_tokenize None-document guard in BM25 reranker", merged: 2026-04-26 }
diff --git a/scripts/check-docs.sh b/scripts/check-docs.sh
index ff065f7200..e428c7e951 100755
--- a/scripts/check-docs.sh
+++ b/scripts/check-docs.sh
@@ -45,13 +45,13 @@ failures=0
 
 # ── 1. test count ────────────────────────────────────────────────────────
 step "1/4  test count in README"
-readme_count=$(grep -oE '^[0-9]+ tests pass on `main`' README.md | grep -oE '^[0-9]+' || echo "")
+readme_count=$(grep -oE '[0-9]+ tests pass on `main`' README.md | grep -oE '^[0-9]+' || echo "")
 if [ -z "$readme_count" ]; then
-    warn "README has no '<N> tests pass on \`main\`' line — skipping"
+    warn "README has no '<N> tests pass on \`main\`' phrase — skipping"
 else
     # Prefer the repo venv's pytest so the check works without an
     # activated environment. Falls back to whatever pytest is on PATH.
-    pytest_bin="$REPO_ROOT/venv/bin/pytest"
+    pytest_bin="$REPO_ROOT/.venv/bin/pytest"
     [ -x "$pytest_bin" ] || pytest_bin="$(command -v pytest 2>/dev/null || true)"
     if [ -z "$pytest_bin" ]; then
         warn "no pytest available — skipping"
@@ -79,7 +79,7 @@ step "2/4  commit hashes referenced in docs resolve"
 docs=(README.md CLAUDE.md FORK_CHANGELOG.md)
 # Strip cross-repo URLs first so we only check hashes that should resolve
 # in *this* fork. Pattern: anything inside (https://github.com/<other>/<repo>/commit/HASH)
-# where <other>/<repo> is not jphein/mempalace.
+# where <other>/<repo> is not techempower-org/mempalace.
 # For each line, skip the line entirely if it mentions a sibling repo
 # (palace-daemon / multipass-structural-memory-eval) — we can't tell which
 # hashes on that line are fork-mempalace vs cross-repo without parsing
@@ -88,7 +88,7 @@ docs=(README.md CLAUDE.md FORK_CHANGELOG.md)
 # adjacent to a sibling-repo mention) but no false positives.
 mapfile -t hashes < <(
     for d in "${docs[@]}"; do
-        grep -v -E 'palace-daemon|multipass-structural-memory-eval|/jphein/[a-z-]+/commit/' "$d" 2>/dev/null
+        grep -v -E 'palace-daemon|multipass-structural-memory-eval|/(jphein|techempower-org)/[a-z-]+/commit/' "$d" 2>/dev/null
     done | grep -hoE '`[0-9a-f]{7,40}`' | tr -d '`' | sort -u
 )
 unresolved=0
@@ -106,7 +106,7 @@ fi
 step "3/4  FORK_CHANGELOG.md regenerates clean"
 render_bin="$REPO_ROOT/scripts/render-docs.py"
 if [ -x "$render_bin" ]; then
-    py="$REPO_ROOT/venv/bin/python"
+    py="$REPO_ROOT/.venv/bin/python"
     [ -x "$py" ] || py="$(command -v python3 2>/dev/null || true)"
     if [ -z "$py" ]; then
         warn "no python interpreter — skipping render check"
@@ -201,7 +201,7 @@ fi
 
 # Check 5 (fork-only YAML commits → CLAUDE.md row inventory) retired
 # 2026-05-11: the CLAUDE.md row inventory it validated was removed in
-# favor of a pointer block to FORK_CHANGELOG.md + jphein/mempalace
+# favor of a pointer block to FORK_CHANGELOG.md + techempower-org/mempalace
 # issues. Check 3 (FORK_CHANGELOG.md ↔ YAML) already guarantees the
 # meaningful sync property — every fork-only YAML commit appears in the
 # rendered FORK_CHANGELOG.md by construction, so a separate CLAUDE.md
diff --git a/scripts/preflight.sh b/scripts/preflight.sh
index 9e6e11cfe3..6888fa1ba7 100755
--- a/scripts/preflight.sh
+++ b/scripts/preflight.sh
@@ -11,9 +11,9 @@
 set -e
 cd "$(dirname "$0")/.."
 
-if [ -x venv/bin/python ]; then
+if [ -x .venv/bin/python ]; then
   # shellcheck disable=SC1091
-  source venv/bin/activate
+  source .venv/bin/activate
 fi
 
 echo "→ ruff check ."
diff --git a/scripts/render-docs.py b/scripts/render-docs.py
index a6f4df53a4..e5faacf261 100755
--- a/scripts/render-docs.py
+++ b/scripts/render-docs.py
@@ -39,7 +39,7 @@
 
 CHANGELOG_PATH = REPO_ROOT / "FORK_CHANGELOG.md"
 CHANGELOG_HEADER = """\
-# Fork Changelog (jphein/mempalace)
+# Fork Changelog (techempower-org/mempalace)
 
 Fork-ahead changes that aren't yet in upstream `MemPalace/mempalace`.
 Upstream's release history lives in [`CHANGELOG.md`](CHANGELOG.md);
@@ -74,7 +74,7 @@ def load_manifest(path: Path = YAML_PATH) -> dict:
 
 def commit_link(sha: str) -> str:
     """Render a 7-char SHA as a markdown link to the fork commit."""
-    return f"[`{sha}`](https://github.com/jphein/mempalace/commit/{sha})"
+    return f"[`{sha}`](https://github.com/techempower-org/mempalace/commit/{sha})"
 
 
 def render_entry(entry: dict[str, Any]) -> str: