TEST: Make score observation contention checks deterministic

[Richard Lundeen (richlundeen)]

Description

Make the score-observation insert/delete contention test independent of the time needed to commit the holder transaction.

The Windows Python 3.13 dev job for #3195 failed in [True-False] with sqlite3.OperationalError: database is locked at the competing BEGIN IMMEDIATE. The test call took 25.23 seconds. The event used by the test fires before SQLite attempts to acquire the lock. It does not guarantee that the holder can commit before SQLite's busy timeout expires. The log does not establish the exact cause of that delay.

A local probe reproduced the same error by holding the first transaction open until the competing write returned SQLITE_BUSY. The probe used a 1 ms busy timeout to avoid a slow reproduction.

The revised test:

• Keeps the holder transaction open while a second session makes a real write attempt.
• Sets busy_timeout=0 on competing connections and requires the exact SQLITE_BUSY error at BEGIN IMMEDIATE. Other database errors still fail the test.
• Commits the holder, then makes the competing operation in a fresh session. The deletion-first case must reject the missing observation; the insertion-first case must retain the replay score's observation link.
• Retains the observation and score-count assertions and adds an exact score-link assertion.

This is a test-only change. It does not add production retries, change production timeouts, or change memory responsibilities. It is independent of the analytics changes in #3195.

Tests and Documentation

Local validation used Windows and uv-managed Python 3.14.2:

• uv sync --frozen --extra all: passed; uv.lock is unchanged.
• uv run --no-sync pytest tests\unit\memory\memory_interface\test_interface_observations.py -q --tb=short: passed, 48 tests.
• uv run --no-sync pytest tests\unit\memory\memory_interface\test_interface_observations.py -k concurrent_insert_and_final_delete -q --tb=short: passed in 10 separate runs, 40 cases total.
• uv run --no-sync pre-commit run --all-files: passed, all hooks. The first run formatted the new assertion; the second run passed without changes.

No public behavior or documentation changed. JupyText execution and the full unit suite were not run.