Give the write path a CLI and an MCP face

remember / supersede / forget now reach the one Python writer from a
terminal, from MCP, and from HTTP. The server, not the caller, decides
who owns a memory.

- cli: remember, supersede, forget. --holder falls back to
  settings.default_holder until the holder registry lands. Timestamps are
  parsed once with one clear message.
- ingestion: import pymupdf as fitz. The deprecated fitz import printed a
  warning to stdout, so every CLI command's JSON answer was unparseable.
- mcp: remember, supersede_memory and forget_memory forward to the Python
  writer over HTTP (ELY_EYE_API, default http://127.0.0.1:8765). SQLite
  stays read-only to the Go side, so the write rules live in one place.
- mcp reads follow schema v2: atoms.time became ingested_at, and reads go
  through v_atoms_current, so a superseded memory stops being findable.
  holder, origin, scenario and event_time now show up in fetch_atom and
  search_atoms.
- scripts/mcp_write_smoke.sh proves the loop end to end: remember, search,
  supersede, forget, then not found.

Evidence: 92 pytest passed; go build, go vet and go test clean; mypy clean
on the 10 touched source files with a fresh cache; ruff clean except a
pre-existing B008 at cli.py:292, identical at HEAD.

backend/src/ely_eye/memory.py

@@ -0,0 +1,219 @@
+"""The single writer for memory: remember, supersede, forget.
+
+Ingestion turns artefacts into atoms. This module turns statements into memories, and it
+is the only place a memory is created or destroyed, so provenance is stamped here and no
+caller can claim to be somebody else.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+from datetime import datetime
+from pathlib import Path
+from typing import Any
+
+from .config import Settings, get_settings
+from .db import Database
+from .schemas import (
+    EvidenceAtom,
+    ForgetReport,
+    MemoryOrigin,
+    Modality,
+    SourceRecord,
+    TimePrecision,
+    TrustScores,
+    Visibility,
+    utc_now,
+)
+from .storage import token_equivalent
+
+# Trust scores describe extraction fidelity, not truthfulness: a remembered sentence is
+# stored verbatim, so the parser is fully trusted and origin carries the epistemic weight.
+MEMORY_TRUST = TrustScores(parser=1.0, ocr=0.0, model=0.0)
+
+
+class MemoryService:
+    def __init__(self, settings: Settings | None = None, database: Database | None = None) -> None:
+        self.settings = settings or get_settings()
+        self.db = database or Database(self.settings)
+
+    def remember(
+        self,
+        text: str,
+        *,
+        holder_id: str,
+        origin: MemoryOrigin,
+        scenario: str = "",
+        event_time: datetime | None = None,
+        event_time_precision: TimePrecision = TimePrecision.unknown,
+        event_time_source: str = "",
+        visibility: Visibility | None = None,
+        supersedes: list[str] | None = None,
+        valid_from: datetime | None = None,
+        retention_until: datetime | None = None,
+    ) -> EvidenceAtom:
+        """Store one statement as a memory owned by a holder.
+
+        Repeating the same sentence as the same holder updates the memory instead of
+        duplicating it; two holders holding the same sentence keep two memories.
+        """
+        statement = text.strip()
+        if not statement:
+            raise ValueError("remember needs text; an empty memory will be found by nobody")
+        if not holder_id.strip():
+            raise ValueError("remember needs a holder; a memory nobody owns cannot be referred to")
+        if event_time is not None and event_time_precision is TimePrecision.unknown:
+            raise ValueError(
+                f"event_time {event_time.isoformat()} needs its precision; "
+                "a day stored as an instant will be read as a moment"
+            )
+
+        moment = utc_now()
+        digest = hashlib.sha256(f"{holder_id}\n{statement}".encode()).hexdigest()
+        source_id = f"mem-{digest[:16]}"
+        self.db.upsert_sources(
+            [
+                SourceRecord(
+                    source_id=source_id,
+                    path=f"memory://{holder_id}/{scenario or 'general'}",
+                    kind="memory",
+                    mime="text/plain",
+                    sha256=hashlib.sha256(statement.encode("utf-8")).hexdigest(),
+                    size_bytes=len(statement.encode("utf-8")),
+                    created_at=moment,
+                    metadata={"holder_id": holder_id, "origin": origin.value},
+                )
+            ]
+        )
+        atom = EvidenceAtom(
+            atom_id=f"{source_id}:text",
+            modality=Modality.text,
+            source_id=source_id,
+            source=f"memory://{holder_id}/{scenario or 'general'}",
+            ingested_at=moment,
+            text=statement,
+            trust=MEMORY_TRUST,
+            token_equivalent=token_equivalent(statement),
+            holder_id=holder_id,
+            origin=origin,
+            scenario=scenario,
+            event_time=event_time,
+            event_time_precision=event_time_precision,
+            event_time_source=event_time_source,
+            supersedes=list(supersedes or []),
+            valid_from=valid_from,
+            visibility=visibility or Visibility(),
+            retention_until=retention_until,
+        )
+        self.db.upsert_atoms([atom])
+        return atom
+
+    def supersede(
+        self,
+        atom_id: str,
+        *,
+        text: str,
+        holder_id: str,
+        origin: MemoryOrigin = MemoryOrigin.self_,
+        scenario: str = "",
+        event_time: datetime | None = None,
+        event_time_precision: TimePrecision = TimePrecision.unknown,
+        event_time_source: str = "",
+    ) -> EvidenceAtom:
+        """Close an old memory and record the new one as its replacement.
+
+        Both rows stay: the new atom answers what is true now, the closed one answers what
+        was true when asked, and a system that forgets the second cannot explain itself.
+        """
+        previous = self.db.get_atoms([atom_id])
+        if not previous:
+            raise ValueError(f"cannot supersede {atom_id}: no such atom")
+        if previous[0].valid_to is not None:
+            raise ValueError(
+                f"{atom_id} was already closed at {previous[0].valid_to.isoformat()}; "
+                "superseding a closed memory hides which replacement is current"
+            )
+        # A correction takes effect when it is asserted: the old memory stops being current
+        # and the new one starts there, so the pair never claims to be true at the same time.
+        moment = event_time or utc_now()
+        replacement = self.remember(
+            text,
+            holder_id=holder_id,
+            origin=origin,
+            scenario=scenario or previous[0].scenario,
+            event_time=event_time,
+            event_time_precision=event_time_precision,
+            event_time_source=event_time_source,
+            supersedes=[atom_id],
+            valid_from=moment,
+        )
+        self.db.close_atoms([atom_id], moment)
+        return replacement
+
+    def forget(
+        self,
+        *,
+        atom_id: str | None = None,
+        source_id: str | None = None,
+        reason: str = "",
+    ) -> ForgetReport:
+        """Delete a memory and everything derived from it.
+
+        The cascade is the whole point: an answer cached from a forgotten memory is still
+        on screen, and a vector index that still carries it will keep surfacing it.
+        """
+        if (atom_id is None) == (source_id is None):
+            raise ValueError("forget takes exactly one of atom_id or source_id")
+        target_source = source_id or self.db.source_id_of_atom(str(atom_id))
+        if target_source is None:
+            raise ValueError(f"nothing to forget: no atom {atom_id}")
+        doomed = self.db.atoms_of_source(target_source)
+        if not doomed:
+            raise ValueError(f"nothing to forget: source {target_source} holds no atom")
+
+        doomed_ids = [atom.atom_id for atom in doomed]
+        edges_deleted = self.db.delete_edges_touching(doomed_ids)
+        atoms_deleted = self.db.delete_atoms(doomed_ids)
+        cache_cleared = self.db.clear_cache_events(f"forget:{target_source}")
+        vectors_dropped = self._drop_vector_index()
+        return ForgetReport(
+            source_id=target_source,
+            atom_ids=doomed_ids,
+            atoms_deleted=atoms_deleted,
+            edges_deleted=edges_deleted,
+            cache_entries_cleared=cache_cleared,
+            vector_index_dropped=vectors_dropped,
+            affected_cartridges=self._cartridges_holding(target_source),
+            reason=reason,
+        )
+
+    def _drop_vector_index(self) -> bool:
+        """The dense index is derived data; a stale one keeps answering with deleted text."""
+        dropped = False
+        for path in (self.settings.cache_dir / "dense_index.npy", self.settings.cache_dir / "dense_index_atoms.json"):
+            if path.exists():
+                path.unlink()
+                dropped = True
+        return dropped
+
+    def _cartridges_holding(self, source_id: str) -> list[dict[str, Any]]:
+        """Name the cartridges that exported this source, so they can be rebuilt.
+
+        An exported cartridge is an immutable artefact on disk; rewriting it here would
+        silently change a DNA hash somebody already quoted.
+        """
+        affected: list[dict[str, Any]] = []
+        for cartridge in self.db.list_cartridges():
+            sources_path = Path(str(cartridge["root_path"])) / "sources.json"
+            try:
+                payload = json.loads(sources_path.read_text(encoding="utf-8"))
+            except FileNotFoundError:
+                affected.append({"cartridge_id": cartridge["cartridge_id"], "contains_source": "unknown"})
+                continue
+            except (OSError, json.JSONDecodeError) as error:
+                raise ValueError(f"{sources_path}: the cartridge source list is unreadable ({error})") from error
+            held = any(str(row.get("source_id")) == source_id for row in payload)
+            if held:
+                affected.append({"cartridge_id": cartridge["cartridge_id"], "contains_source": True})
+        return affected