fix(agent): rehydrate session accumulators from SQLite on restart (#67762)

[JonthanaHanh]

Problem

When the gateway restarts mid-session, init_agent() resets all session accumulators (session_estimated_cost_usd, token counts, session_api_calls) to zero. The persisted data in SQLite (sessions.estimated_cost_usd, session_model_usage rows) is correct, but nothing copies it back into the agent's in-memory state.

This means:

• The live cost counter shows only post-restart costs
• Token totals reset to zero
• /insights dashboard shows the correct total, but the agent's runtime state lies

Verified against main (commit 26480e6c5).

Reproduction

1. Start a session, run several turns (cost accumulates to, say, $4.27).
2. Restart the gateway process.
3. Resume the session by sending a new message.
4. Inspect agent.session_estimated_cost_usd -- shows $0.0 (or only post-restart cost).
5. Check sessions.estimated_cost_usd in ~/.hermes/state.db -- correctly shows the full total.

Fix

After the accumulator reset block in agent/agent_init.py, read the existing session row from SQLite (if present) and restore the persisted totals:

• session_estimated_cost_usd
• session_cost_status / session_cost_source
• All token counters (input, output, cache, reasoning)
• session_api_calls

For new sessions, get_session() returns None and the block is a no-op. DB errors are caught and swallowed (fail-open).

Files Changed

• agent/agent_init.py -- add rehydration block after accumulator reset

Fixes #67762

[alt-glitch]

This was generated by AI during triage.

Related to #67770 and #67790. This PR restores additional direct session-row accumulators, while #67770 uses tested per-model cost aggregation and #67790 addresses sticky status semantics; please choose or consolidate the overlapping rehydration work.

[teknium1]

Thanks for targeting the live-accounting reset; current main still resets these fields at agent/agent_init.py:2522-2533.

Problems

• The proposed get_session() source is incomplete for this purpose. record_auxiliary_usage() intentionally writes only to session_model_usage, not sessions (hermes_state.py:4920-4930), so this restores an undercount after any auxiliary call.
• agent/conversation_loop.py:2863-2871 maintains prompt/completion/total separately from canonical input/output/cache/reasoning totals. Mapping persisted input_tokens and output_tokens into both counter families changes their meaning after restart.
• Commit 1a3d88d changes no tests, so the restart and source-divergence contract is unpinned.

Suggested changes

• Aggregate the durable session_model_usage rows for the rehydrated fields, and preserve the distinction between canonical and display token counters.
• Add a restart regression test using the same temporary state database, including an auxiliary-usage row.

The reset hunk is still present on current main, but the accounting source and counter mapping need rework. This is an automated hermes-sweeper review.