superlocalmemory 4.1.3 → 4.1.5

package/src/superlocalmemory/cli/diagnostics_cmd.py

@@ -10,8 +10,11 @@ from argparse import Namespace
 
 def cmd_diagnostics(args: Namespace) -> None:
     action = getattr(args, "diagnostics_command", None)
+    if action == "reliability":
+        _cmd_reliability(args)
+        return
     if action != "export":
-        raise SystemExit("choose a diagnostics subcommand: export")
+        raise SystemExit("choose a diagnostics subcommand: export, reliability")
 
     from superlocalmemory.cli.json_output import json_print
     from superlocalmemory.infra.local_diagnostics import default_diagnostics
@@ -25,4 +28,74 @@ def cmd_diagnostics(args: Namespace) -> None:
     print("No automatic reporting was enabled.")
 
 
+def _cmd_reliability(args: Namespace) -> None:
+    """Report whether wired mechanisms are effective, not merely present.
+
+    ``implemented``, ``reachable`` and ``effective`` are three different
+    questions. A grep answers the first and an import answers the second; only
+    querying this store answers the third. Both checks are read-only.
+    """
+    from superlocalmemory.cli.json_output import json_print
+    from superlocalmemory.infra.data_root import state_path
+    from superlocalmemory.reliability import (
+        DEFAULT_MIN_OBSERVATIONS,
+        check_beta_learners,
+        check_schema_guards,
+    )
+
+    floor = getattr(args, "min_observations", None) or DEFAULT_MIN_OBSERVATIONS
+    learners = check_beta_learners(state_path("learning.db"), min_observations=floor)
+    guards = check_schema_guards(state_path("memory.db"))
+
+    payload = {
+        "learners": [
+            {
+                "table": v.table,
+                "verdict": v.verdict,
+                "units": v.units,
+                "units_at_prior_mean": v.units_at_prior_mean,
+                "units_matching_neutral_identity": v.units_matching_neutral_identity,
+                "observations": v.observations,
+                "detail": v.detail,
+            }
+            for v in learners
+        ],
+        "schema_guards": [
+            {
+                "name": g.name,
+                "verdict": g.verdict,
+                "missing": list(g.missing),
+                "found_elsewhere": [
+                    {
+                        "table": tbl,
+                        "column": col,
+                        "populated_rows": n,
+                        # -1 when coverage over the guarded table could not be
+                        # measured. Populated rows alone overstate the remedy.
+                        "coverage_pct_of_guarded_table": cov,
+                    }
+                    for tbl, col, n, cov in g.found_elsewhere
+                ],
+                "detail": g.detail,
+            }
+            for g in guards
+        ],
+    }
+
+    if bool(getattr(args, "json", False)):
+        json_print("diagnostics-reliability", data=payload)
+        return
+
+    if not learners and not guards:
+        print("No Bayesian learners or schema-guarded paths found in this store.")
+        return
+
+    for v in learners:
+        print(f"[{v.verdict}] {v.table}")
+        print(f"    {v.detail}")
+    for g in guards:
+        print(f"[{g.verdict}] {g.name}")
+        print(f"    {g.detail}")
+
+
 __all__ = ["cmd_diagnostics"]

package/src/superlocalmemory/cli/main.py

@@ -875,6 +875,18 @@ def main() -> None:
         "export", help="Write a deterministic content-free JSON report",
     )
     diagnostics_export.add_argument("destination")
+    diagnostics_reliability = diagnostics_sub.add_parser(
+        "reliability",
+        help=(
+            "Ask whether wired mechanisms are actually effective: has each "
+            "Bayesian learner moved off its prior, and has each schema-guarded "
+            "path ever executed against this store"
+        ),
+    )
+    diagnostics_reliability.add_argument(
+        "--min-observations", type=int, default=None,
+        help="Observation floor below which an unmoved posterior is not reported",
+    )
     diagnostics_export.add_argument(
         "--json", action="store_true", help="Output structured JSON",
     )

package/src/superlocalmemory/core/remember_runtime.py

@@ -419,8 +419,19 @@ class CanonicalRememberRuntime:
             OwnershipRequiredError,
             WriteCoordinatorError,
         ) as exc:
+            # Name which of the three it was. They are not interchangeable and
+            # they call for different responses: an unavailable journal is I/O
+            # or lock contention and worth retrying, lost ownership means
+            # another writer holds the lease, and a coordinator error is the
+            # same type the generation fence raises to reject a stale epoch.
+            # Collapsing all three into one string makes a spurious fence
+            # rejection indistinguishable from a transient disk stall, for the
+            # operator reading a log and for a caller deciding whether to
+            # retry. Only the class name is included: it is the whole of the
+            # discriminating information and carries no request content.
             raise CanonicalRememberUnavailable(
-                "canonical remember is temporarily unavailable"
+                "canonical remember is temporarily unavailable "
+                f"({type(exc).__name__})"
             ) from exc
         finally:
             clear_admission_epoch(request.profile_id, request.idempotency_key)

package/src/superlocalmemory/reliability/__init__.py

@@ -0,0 +1,45 @@
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under AGPL-3.0-or-later - see LICENSE file
+# Part of SuperLocalMemory | https://qualixar.com | https://varunpratap.com
+
+"""Checks that ask whether a mechanism is *effective*, not merely present.
+
+Three questions get confused in a system this size:
+
+* **Implemented** — the code exists. A grep answers this.
+* **Reachable** — something calls it. A call-graph trace answers this.
+* **Effective** — it has actually changed an outcome against real data.
+  **Neither of the above answers this.** Only querying the store does.
+
+A mechanism can pass the first two questions for months and fail the third
+silently: a learner whose reward channel emits a constant still records plays,
+and a conditional path guarded by a missing column still appears in coverage.
+Nothing raises, nothing logs, and every file is present.
+
+The two checks here answer the third question directly.
+
+* :mod:`.prior_distance` — has a Bayesian learner's posterior actually moved
+  away from its prior?
+* :mod:`.join_liveness` — has a schema-guarded code path ever executed against
+  this store, and if not, which requirement is missing?
+
+Both are read-only, both are fail-soft, and neither is on a hot path.
+"""
+
+from superlocalmemory.reliability.join_liveness import (
+    GuardVerdict,
+    check_schema_guards,
+)
+from superlocalmemory.reliability.prior_distance import (
+    DEFAULT_MIN_OBSERVATIONS,
+    LearnerVerdict,
+    check_beta_learners,
+)
+
+__all__ = [
+    "DEFAULT_MIN_OBSERVATIONS",
+    "GuardVerdict",
+    "LearnerVerdict",
+    "check_beta_learners",
+    "check_schema_guards",
+]

package/src/superlocalmemory/reliability/join_liveness.py

@@ -0,0 +1,301 @@
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under AGPL-3.0-or-later - see LICENSE file
+# Part of SuperLocalMemory | https://qualixar.com | https://varunpratap.com
+
+"""Can a schema-guarded code path run against this store, and has it had the data?
+
+A note on what this proves, because the honest scope is narrower than the
+convenient phrasing. The check reads the schema: it establishes whether the
+guard's requirements are present *now*, and therefore whether the path is
+reachable at all. It does not read an execution history, so a satisfied guard
+means "this would execute", not "this has executed". The inference to "never
+executed" is sound only in the other direction -- if a required column is absent
+from a store that has been in service, the guarded path cannot have run against
+it -- and that is the direction the findings rely on.
+
+Some features are wired behind a guard that asks the schema a question before
+doing any work — "is this table here, does that column exist" — and fall back
+silently when the answer is no. The fallback is correct behaviour: it is what
+keeps an old store openable. But it means a feature can be implemented, called
+on the hot path, and covered by tests, while never once executing against real
+data. Static analysis passes. A call-graph trace passes. Coverage passes. The
+guard returns ``False`` and the feature is arithmetically absent.
+
+Two things make that failure hard to see from inside the code:
+
+1. **The guard is doing its job.** There is no error to raise. Falling back is
+   the designed response to a missing column.
+2. **The fallback is often a neutral value**, which composes into an identity.
+   A decay-rate multiplier that falls back to a trust of 1.0 collapses
+   ``lambda * (1 + kappa * (1 - trust))`` to ``lambda`` — the feature is on, and
+   it computes exactly what having no feature would compute.
+
+So this check does two things a plain schema assertion does not. It records
+whether each named guard **passes right now**, and when a guard fails it looks
+for the required data **elsewhere in the store** — because the common case is
+not that the data is missing, it is that the guard is asking the wrong table.
+
+A guard reported as ``SATISFIED_ELSEWHERE`` is the most actionable outcome
+available: the feature is one re-keyed join away from working, and no backfill
+or migration is required.
+"""
+
+from __future__ import annotations
+
+import logging
+import sqlite3
+from dataclasses import dataclass, field
+from typing import Any
+
+logger = logging.getLogger("superlocalmemory.reliability.join_liveness")
+
+
+@dataclass(frozen=True)
+class Requirement:
+    """One schema object a guard needs: a table, optionally a column on it."""
+
+    table: str
+    column: str | None = None
+
+    def __str__(self) -> str:
+        return f"{self.table}.{self.column}" if self.column else self.table
+
+
+@dataclass(frozen=True)
+class Guard:
+    """A named conditional path and the schema it requires to execute."""
+
+    name: str
+    describes: str
+    requires: tuple[Requirement, ...]
+    #: What the feature computes when the guard fails, in words. Recording this
+    #: is the difference between "a feature is off" and "a feature is off and
+    #: indistinguishable from not having it".
+    fallback_behaviour: str
+
+
+@dataclass(frozen=True)
+class GuardVerdict:
+    name: str
+    describes: str
+    #: LIVE = requirements present, so the path is reachable now. DEAD /     #: SATISFIED_ELSEWHERE = a requirement is absent, so the path cannot have run
+    #: against this store. LIVE is NOT evidence of past execution.
+    verdict: str
+    missing: tuple[str, ...] = field(default=())
+    found_elsewhere: tuple[tuple[str, str, int, float], ...] = field(default=())
+    detail: str = ""
+
+    @property
+    def is_live(self) -> bool:
+        return self.verdict == "LIVE"
+
+
+#: Guards worth reporting on. Each entry names a real conditional in the code,
+#: so that a reader can go from this list to the line that asks the question.
+GUARDS: tuple[Guard, ...] = (
+    Guard(
+        name="trust_weighted_forgetting",
+        describes=(
+            "learning/forgetting_scheduler.py::_has_trust_tables — gates the "
+            "per-fact trust lookup that modulates the decay rate"
+        ),
+        requires=(
+            Requirement("trust_scores"),
+            Requirement("atomic_facts", "created_by"),
+        ),
+        fallback_behaviour=(
+            "every fact takes trust = 1.0, so lambda_eff collapses to "
+            "lambda_base and the decay rate is identical to no trust weighting"
+        ),
+    ),
+)
+
+
+def _table_exists(conn: sqlite3.Connection, table: str) -> bool:
+    return (
+        conn.execute(
+            "SELECT 1 FROM sqlite_master WHERE type='table' AND name=? LIMIT 1",
+            (table,),
+        ).fetchone()
+        is not None
+    )
+
+
+def _columns(conn: sqlite3.Connection, table: str) -> set[str]:
+    try:
+        return {r[1] for r in conn.execute(f'PRAGMA table_info("{table}")')}
+    except sqlite3.Error:
+        return set()
+
+
+def _satisfied(conn: sqlite3.Connection, requirement: Requirement) -> bool:
+    if not _table_exists(conn, requirement.table):
+        return False
+    if requirement.column is None:
+        return True
+    return requirement.column in _columns(conn, requirement.table)
+
+
+def _look_elsewhere(
+    conn: sqlite3.Connection,
+    column: str,
+    *,
+    join_target: str | None = None,
+    join_key: str = "fact_id",
+) -> list[tuple[str, str, int]]:
+    """Find other tables carrying ``column``, with a populated-row count.
+
+    This is the part that turns a failed guard into a fix. A column the guard
+    could not find on its own table is often present on a neighbouring one,
+    already populated.
+
+    **A populated count is not coverage, and reporting it alone overstates the
+    remedy.** A provenance-style table can carry a value on every one of its own
+    rows while describing only part of the set the join needs: rows can be
+    missing for older entities entirely. When ``join_target`` is given, the
+    coverage fraction over that table is measured and returned, because the
+    honest question is not "does this column exist somewhere" but "how much of
+    what the guard needs would the re-keyed join actually resolve".
+    """
+    out: list[tuple[str, str, int]] = []
+    tables = [
+        r[0]
+        for r in conn.execute(
+            "SELECT name FROM sqlite_master WHERE type='table' "
+            "AND name NOT LIKE 'sqlite_%'",
+        )
+    ]
+    for table in tables:
+        if column not in _columns(conn, table):
+            continue
+        try:
+            populated = conn.execute(
+                f'SELECT COUNT(*) FROM "{table}" '
+                f'WHERE "{column}" IS NOT NULL AND TRIM("{column}") <> \'\'',
+            ).fetchone()[0]
+        except sqlite3.Error:
+            populated = 0
+        covered = -1
+        if join_target and join_target != table:
+            try:
+                total = conn.execute(
+                    f'SELECT COUNT(*) FROM "{join_target}"',
+                ).fetchone()[0]
+                if total:
+                    hit = conn.execute(
+                        f'SELECT COUNT(*) FROM "{join_target}" t WHERE EXISTS ('
+                        f'  SELECT 1 FROM "{table}" s WHERE s."{join_key}" = t."{join_key}"'
+                        f'  AND s."{column}" IS NOT NULL AND TRIM(s."{column}") <> \'\')',
+                    ).fetchone()[0]
+                    covered = round(100.0 * hit / total, 1)
+            except sqlite3.Error:
+                covered = -1
+        out.append((table, column, int(populated), covered))
+    return out
+
+
+def _evaluate(conn: sqlite3.Connection, guard: Guard) -> GuardVerdict:
+    missing = [str(r) for r in guard.requires if not _satisfied(conn, r)]
+    if not missing:
+        return GuardVerdict(
+            name=guard.name,
+            describes=guard.describes,
+            verdict="LIVE",
+            detail=(
+                "Every requirement is present, so the guarded path is reachable "
+                "on this store. This is schema evidence, not execution history: "
+                "it does not establish that the path has ever actually run."
+            ),
+        )
+
+    elsewhere: list[tuple[str, str, int, float]] = []
+    for requirement in guard.requires:
+        if requirement.column and not _satisfied(conn, requirement):
+            for found in _look_elsewhere(
+                conn, requirement.column, join_target=requirement.table,
+            ):
+                if found[0] != requirement.table:
+                    elsewhere.append(found)
+
+    populated = [e for e in elsewhere if e[2] > 0]
+    if populated:
+        # Rank by coverage of the guarded table, falling back to row count only
+        # when coverage could not be measured. Picking the largest table instead
+        # would recommend a 4,000-row partial source over a smaller complete one.
+        best = max(populated, key=lambda e: (e[3], e[2]))
+        cov = best[3]
+        if cov < 0:
+            remedy = (
+                f"Re-keying the join onto that table would make the path "
+                f"executable; coverage over the guarded table could not be "
+                f"measured here, so confirm it before relying on the remedy."
+            )
+        elif cov >= 99.5:
+            remedy = (
+                f"That table covers {cov}% of the rows the guard needs, so "
+                f"re-keying the join makes the path live without a backfill."
+            )
+        else:
+            remedy = (
+                f"That table covers only {cov}% of the rows the guard needs. "
+                f"Re-keying the join makes the path executable for those rows "
+                f"and leaves the remainder on the same fallback, so this is a "
+                f"partial remedy and a backfill decision, not a free fix."
+            )
+        return GuardVerdict(
+            name=guard.name,
+            describes=guard.describes,
+            verdict="SATISFIED_ELSEWHERE",
+            missing=tuple(missing),
+            found_elsewhere=tuple(elsewhere),
+            detail=(
+                f"The guard requires {', '.join(missing)}, which is absent, so "
+                f"the path has never executed against this store — "
+                f"{guard.fallback_behaviour}. The data it needs is present on "
+                f"{best[0]}.{best[1]} with {best[2]} populated rows. {remedy}"
+            ),
+        )
+
+    return GuardVerdict(
+        name=guard.name,
+        describes=guard.describes,
+        verdict="DEAD",
+        missing=tuple(missing),
+        detail=(
+            f"The guard requires {', '.join(missing)}, which is absent from "
+            f"this store and not carried by any other table, so the path has "
+            f"never executed — {guard.fallback_behaviour}."
+        ),
+    )
+
+
+def check_schema_guards(
+    memory_db: Any, *, guards: tuple[Guard, ...] = GUARDS,
+) -> list[GuardVerdict]:
+    """Report, per registered guard, whether its path can run on this store.
+
+    ``memory_db`` may be a path or an open connection. Read-only, and fail-soft:
+    an error yields an empty list, because a diagnostic must never be the reason
+    something breaks.
+    """
+    owns_connection = not isinstance(memory_db, sqlite3.Connection)
+    conn: sqlite3.Connection | None = None
+    try:
+        conn = (
+            sqlite3.connect(f"file:{memory_db}?mode=ro", uri=True)
+            if owns_connection
+            else memory_db
+        )
+        return [_evaluate(conn, guard) for guard in guards]
+    except Exception:
+        logger.debug("join-liveness check skipped", exc_info=True)
+        return []
+    finally:
+        if owns_connection and conn is not None:
+            try:
+                conn.close()
+            except Exception:
+                pass
+
+
+__all__ = ["GUARDS", "Guard", "GuardVerdict", "Requirement", "check_schema_guards"]