v0.11.0 — attachable memory

Attachable memory

A pack is a sealed, signed YantrikDB file that a host database mounts to gain
the knowledge and rules inside — and unmounts to give them back, leaving the host
byte-for-byte as it was.

from yantrikdb import YantrikDB

db = YantrikDB("memory.db", 64)
db.mount_pack("wordpress-expert-0.2.0.ydbpack")   # transient
db.install_pack("wordpress-expert-0.2.0.ydbpack") # survives restart
db.pack_context()                                  # rules to inject
db.unmount_pack(pack_id)

CLI: yantrikdb pack info|install|list|uninstall|keygen|sign.

This is the release the pack marketplace has been
waiting on. Until now, pip install yantrikdb gave you a build with no
mount_pack at all, so every listing in that catalog was unmountable by anyone
who downloaded it.

Mount is not import

Importing copies rows into the host: it leaves tombstones on removal and
permanently shifts namespace_importance_stats.count, so the host can never be
returned to its prior state. Mounting adds a read-only pool alongside the host's,
merges at recall time, and drops the handle on unmount.

unmount_leaves_host_byte_identical is the test that holds this, and it compares
file bytes rather than row counts — a row count would not have caught the stats
drift, which is precisely why import was the wrong verb.

Three tiers, because knowledge and ability are different mechanisms

tier	delivery	what it is
constitution	injected unconditionally	ability
corpus	retrieved by similarity	knowledge
coverage	advertised topics	what the pack claims

Measured rather than assumed. Across eight ability tasks on three models,
retrieval alone scored +1 / +1 / +0; the constitution scored +5 / +4 / +5.
Seven of the eight tasks retrieved zero facts — a rule the model must apply is
not a rule it thinks to look up.

Injection is gated on similarity, not the composite score

Ungated top-k injection measured 12/12 → 5/12 on an unrelated control set.
The content was not wrong; the model concluded it was only permitted to answer
from the supplied material and began refusing questions like "what is the capital
of Japan". A similarity floor of 0.55 restores 12/12 with no loss on the pack's
own questions.

It has to be similarity rather than the composite recall score, because composite
folds in importance and recency — near-uniform across a freshly built pack, and
therefore carrying no relevance signal:

	top-1 similarity	top-1 composite
on-topic	0.65 – 0.79	0.83 – 0.91
off-topic	0.09 – 0.45	0.37 – 0.69

Trust

Ed25519 over a length-prefixed payload covering origin, name, version, namespace,
content digest, embedder identity, row count, and every constitution and coverage
line. description is deliberately excluded, so a store can localise a listing
without invalidating the signature. Trust-on-first-use, like SSH.

Fixes

• Embedder identity was never persisted. SearchState::initial reset it to
  ExternalOrUnknown on every open, so the same-dimension-different-model guard
  was unreachable and no database in the wild had provable embedder identity.
  BundledEmbedder also had no fingerprint() despite the trait doc promising
  one. Identity now persists to meta.
• recall() short-circuited on an empty host index, which broke the flagship
  case of mounting a pack into a fresh database.
• mcp extra bounded below 2.0.0. It was declared mcp>=1.0.0 with no
  ceiling, so the day the SDK published 2.0.0 — removing mcp.server.fastmcp —
  CI went red on every platform with no commit in between.

Compatibility

No schema migration for existing databases; pack state lives in new tables and
meta keys added on open. Nothing in the recall path changes for a database with
no packs mounted.

Full changelog: v0.10.1...v0.11.0