feat(creds): cred-reuse foundation + vectorstore scaffold

Lays the storage and bus substrate for the "credential reuse patterns"
task in DEVELOPMENT.md and scaffolds decnet/vectorstore/ as the future
substrate for statistical attacker re-identification over behavioral
fingerprints. No correlator, profiler, API, or dashboard wiring in
this commit — see TODO.md for the handoff.

Schema:
  - Credential.attacker_uuid (nullable FK to attackers.uuid),
    backfilled by the profiler post-write to avoid coupling the
    capture path to the profiler's ordering.
  - CredentialReuse table — UUID PK, JSON list columns for the
    accumulating attacker_uuids/ips/deckies/services, target_count
    (the discriminative scalar), confidence reserved for a future
    fuzzy-credential pass.

Repo:
  - upsert_credential_reuse / list_credential_reuses /
    get_credential_reuse_by_id / update_credential_attacker_uuid.
  - Renamed pre-existing get_credential_reuse(secret_sha256) to
    get_credential_attempts_for_secret(secret_sha256) — the new
    findings table needs the cleaner name.

Bus topics:
  - credential.captured (one per Credential upsert)
  - credential.reuse.detected (correlator-emitted on insert/grow)

Vectorstore subpackage (decnet/vectorstore/, flat layout mirroring
decnet/bus/):
  - BaseVectorStore ABC keyed by (kind, id) — kind discriminator
    means new feature families are additive, no schema migration.
  - FakeVectorStore (in-memory L2 KNN), NullVectorStore (no-op for
    DECNET_VECTORSTORE_ENABLED=false), SqliteVecVectorStore (lazy
    sqlite_vec extension load, one vec0 virtual table per kind).
  - get_vectorstore() env-driven dispatch with graceful fallback
    to FakeVectorStore when the sqlite-vec extension isn't on the
    host, so workers don't crash on a missing optional dep.

Tests: 26 new (11 cred-reuse repo, 15 vectorstore). Existing
credentials and base-repo tests updated for the rename. Total: 34
passing on the touched files.
This commit is contained in:
2026-04-26 03:18:34 -04:00
parent 817ce32e6d
commit ce4be68501
17 changed files with 1615 additions and 11 deletions

View File

@@ -14,6 +14,8 @@ Token structure (NATS-style, dot-separated):
attacker.scored
attacker.session.started
attacker.session.ended
credential.captured
credential.reuse.detected
system.log
system.bus.health
system.{worker}.health
@@ -32,6 +34,7 @@ TOPOLOGY = "topology"
DECKY = "decky"
ATTACKER = "attacker"
SYSTEM = "system"
CREDENTIAL = "credential"
# ─── Leaf event-type constants (the last segment of each topic) ──────────────
@@ -75,6 +78,15 @@ ATTACKER_FINGERPRINTED = "fingerprinted"
ATTACKER_SESSION_STARTED = "session.started"
ATTACKER_SESSION_ENDED = "session.ended"
# Credential event types (second/third tokens under ``credential``).
# ``credential.captured`` fires once per upserted Credential row — the
# correlator listens for it and runs the cred-reuse query in response,
# so reuse detection latency is sub-second after a fresh capture.
# ``credential.reuse.detected`` fires when the correlator inserts a new
# CredentialReuse row or grows an existing one (added decky/service/IP).
CREDENTIAL_CAPTURED = "captured"
CREDENTIAL_REUSE_DETECTED = "reuse.detected"
# System event types.
SYSTEM_LOG = "log"
SYSTEM_BUS_HEALTH = "bus.health"
@@ -143,6 +155,19 @@ def system(event_type: str) -> str:
return f"{SYSTEM}.{event_type}"
def credential(event_type: str) -> str:
"""Build ``credential.<event_type>``.
*event_type* is typically one of :data:`CREDENTIAL_CAPTURED` or
:data:`CREDENTIAL_REUSE_DETECTED`. Dotted leaves
(``reuse.detected``) are permitted — same rationale as
:func:`system`.
"""
if not event_type:
raise ValueError("credential topic requires a non-empty event_type")
return f"{CREDENTIAL}.{event_type}"
def attacker(event_type: str) -> str:
"""Build ``attacker.<event_type>``.