Agents - bot-mode reporting includes stale controller rows from inactive runtimes

[mlguys]

Area

Agents & chat

What happened

A bot-operated Agent reported open positions, exposure, and PnL from inactive
prior bot runtimes as though they were still live.

At one checkpoint, the active-bot endpoint returned exactly one current runtime.
Its controller custom_info schema-3 ledger reported 2 active positions and
0.04115160993802155 quote PnL. Condor reported 24 running positions,
13.43899382355312 quote exposure, and 0.09556344473681769 quote PnL.

The latest controller-performance response contained nine prior runtime records
that were absent from the active-bot endpoint but still carried status=running.
Those records contributed 22 position rows and 8.694458... quote exposure to
the Agent report.

The complete sanitized observation and arithmetic are attached as
condor-agent-reporting-evidence.txt.

Expected

• Current open positions, exposure, and unrealized PnL should include only bot
  instances confirmed by the active-bot inventory.
• Prior-runtime running records should not become live merely because they
  remain in the latest-performance response.
• Historical realized PnL may continue through the existing history path.
• If a controller provides an explicit reporting ledger in custom_info, Condor
  should preserve its provenance instead of silently mixing it with incompatible
  generic metrics.

Likely source path

• resolve_bots() calls every matching entry in all_bot_perf live without reconciling it against the active-bot endpoint.
• bot_executor_rows() converts every merged positions_summary entry into a live-looking row and derives exposure as abs(amount) * breakeven_price.
• ExecutorsProvider keeps rows whose controller status says RUNNING.
• Exposure aggregation sums all of those rows.
• Agent risk-state wiring consumes the resulting exposure.

A narrow fix would reconcile live performance rows against active bot instance
identities before building the current open book. Suggested regression coverage
belongs in tests/test_bot_performance.py and/or
tests/test_session_pnl_attribution.py.

Steps to reproduce

1. Use a stable bot base that has a current runtime and an older runtime.
2. Return only the current runtime from the active-bot endpoint.
3. Return both runtimes from latest controller performance, with both records
   still reporting status=running and carrying positions_summary rows.
4. Fetch the bot-mode Agent performance for the stable base.
5. Observe that positions from both current and inactive runtimes contribute to
   open count, unrealized PnL, and exposure.

Environment

Observed in a private downstream checkout; no private branch or source is
published with this report. The implicated logic was independently confirmed in
public upstream Condor commit
a3997a2b.

Logs

Attach: condor-agent-reporting-evidence.txt

[mlguys]

condor-agent-reporting-evidence.txt

[mlguys]

Closed controllers from other sessions are showing in new session:

Stale executors from other sessions are included in the new session: