fix(memory): coerce legacy float salience on row hydration

[djgaasch-oss]

Problem

Early PinkyBot builds stored reflection salience as a 0–1 float. The Reflection pydantic model now requires int in [1, 5], so a single legacy row raises ValidationError at hydration and poisons every recall() / memory_query that touches it.

Observed live on the bubbs agent (Jul 16): 3 rows with salience 0.7 / 0.85 / 0.9 made semantic recall completely unusable:

Error executing tool recall: 1 validation error for Reflection
salience
  Input should be a valid integer, got a number with a fractional part

Fix

Add ReflectionStore._coerce_salience() and apply it in _row_to_reflection:

• 0–1 floats map onto the 1–5 scale (x*5, rounded)
• stray numerics round and clamp to [1, 5]
• non-numeric garbage falls back to the default (3)

Reads become tolerant; writes are already validated by pydantic, so no new bad data can enter.

Testing

• New TestLegacySalienceCoercion: 9 parametrized coercion cases + an end-to-end test that corrupts a stored row to 0.7 and asserts get() hydrates it as 4 instead of raising.
• Full tests/test_memory_store.py: 120 passed, 1 skipped.

🤖 Opened by Bubbs

[djgaasch-oss]

Pushed a correctness fix to the coercion: the legacy 0-1 → ×5 mapping now applies only to genuine floats. SQLite preserves per-value type (legacy rows hydrate as float, modern rows as int), and without the type check a perfectly valid modern salience of int 1 fell inside 0 <= val <= 1 and was silently inflated to 5 on every read. Bools (int subclass) are treated as garbage → default 3.

New regression tests: (1, 1), (0, 1), (True, 3) param cases + an end-to-end test_modern_salience_one_is_not_inflated. Suite: 124 passed, 1 skipped; ruff clean.

🤖 Opened by Bubbs