feat(#2007): implement mempalace memory_mode kg_backend and replace routing (#2010)

Wire the two forward-declared mempalace.memory_mode modes so they actually
route recall/capture instead of silently behaving as `augment`:

- kg_backend: the palace temporal KG is the primary knowledge-graph source;
  native .planning/graphs/ is the fallback. Non-KG drawer recall stays additive.
- replace: recall resolves through the palace as the source of truth; native
  artifacts are the fallback.

Every mode stays onError:skip and default-resilient — an unreachable palace
degrades to native memory and GSD keeps writing .planning/graphs/, so no memory
is lost. Cross-mode .planning/graphs/ migration remains a documented open
question (PRD/ADR §17), out of scope here.

Surfaces updated (instruction-only contract): recall/capture commands (+ generated
skills), discuss/wave fragments, curator agent, capability.json schema. Docs:
how-to Step 3, CONFIGURATION, FEATURES, CONTEXT glossary. Regenerated
capability-registry, golden install-parity fixtures (mempalace hashes only),
agent-size-baseline. Added a routing-contract + cross-surface parity test.

Incidental (folded per no-defer rule): removed pre-existing unused imports
(spawnSync in capability-registry.test.cjs; fs in issue-498-package-identity.test.cjs)
that eslint flagged in/alongside the touched files.

Closes #2007

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

capabilities/mempalace/capability.json

@@ -37,7 +37,7 @@
         "replace"
       ],
       "default": "augment",
-      "description": "How MemPalace relates to GSD native memory. Only 'augment' (additive) is implemented today; 'kg_backend' and 'replace' are forward-declared (routing seam not yet built) and currently behave as 'augment'."
+      "description": "How MemPalace relates to GSD native memory during recall/capture. 'augment' (default): the palace is an additive layer alongside .planning/graphs/ — native memory stays authoritative. 'kg_backend': knowledge-graph queries resolve against the palace's temporal KG as the primary source, with .planning/graphs/ as fallback. 'replace': recall resolves through the palace as the source of truth, native memory as fallback. Every mode stays onError:skip and default-resilient — an unreachable palace degrades to native memory, and GSD keeps writing .planning/graphs/ so no memory is lost. Cross-mode migration of existing .planning/graphs/ into the palace is a separate, not-yet-implemented concern."
     },
     "mempalace.wing": {
       "type": "string",

capabilities/mempalace/fragments/capture-problems.md

@@ -15,6 +15,6 @@ For each confirmed bug/issue resolved in this wave:
 2. **Dedupe first.** Call `mempalace_check_duplicate` (interactive) before filing so re-runs don't create duplicate drawers.
 3. **File the drawer verbatim.** Store the problem statement and its fix as a drawer in `room: problems` — interactive: `mempalace_add_drawer`; headless: `mempalace mine` / `mempalace hook run`. Include provenance (`source_file`, phase id).
 4. **Mirror the KG fact** when `mempalace.mirror_kg` is on: add `(<bug>, fixed_by, <fix>)` with `valid_from` = the phase date via `mempalace_kg_add`.
-5. **Mode awareness.** Only `augment` is currently wired: the fact is an *additive* mirror alongside `.planning/graphs/` (never a replacement). `kg_backend`/`replace` are forward-declared and behave as `augment` today.
+5. **Mode awareness** (`mempalace.memory_mode`). Under `augment` the fact is an *additive* mirror alongside `.planning/graphs/`. Under `kg_backend`/`replace` the palace is the *authoritative* store for the fact; GSD still writes `.planning/graphs/` through its normal graphify, so an unreachable palace never loses it.
 
 Captures are idempotent: deterministic drawer IDs + `check_duplicate` mean re-running the wave re-files the same content without duplication. On any error, skip and let the wave complete normally.

capabilities/mempalace/fragments/recall-discuss.md

@@ -16,7 +16,7 @@ Otherwise — before gathering new context, surface what you already know. This
 3. **Targeted recall.** Search the palace for prior work on this phase's topic:
    - Interactive → `mempalace_search(query=<phase topic>, wing=<wing>)` and, when `mempalace.mirror_kg` is on, `mempalace_kg_query` / `mempalace_kg_timeline` for decision facts and their validity windows.
    - Headless → `mempalace search "<phase topic>" --wing <wing>`.
-4. **Mode awareness.** Only `augment` is currently wired: always treat the palace as an *additional* recall layer on top of GSD's native memory — never skip `.planning/graphs/` or STATE. `kg_backend`/`replace` are forward-declared and behave as `augment` today.
+4. **Mode awareness** (`mempalace.memory_mode`). `augment` (default): the palace is an *additional* layer — never skip `.planning/graphs/` or STATE. `kg_backend`: query the palace's temporal KG as the *primary* graph source, with `.planning/graphs/` as fallback. `replace`: resolve recall *through the palace as the source of truth*, native memory as fallback. In every mode an unreachable palace falls back to native memory — recall never blocks.
 5. **Surface, don't dump.** Fold the top relevant drawers, decisions, patterns, and *surprises* into the discussion as prior context — cite drawer/fact provenance. Do not paste raw search output.
 
 If any MemPalace call errors or times out, skip the rest of recall and proceed with discussion as normal.