ga: enforce long-term settlement at task completion for 15+ turn tasks

[louisss1016]

Addresses #789.

What

The schema only asks the model to call start_long_term_update when a 15+ turn task completes; the runtime never checks, and the issue's runs finished on turn 20/39 with zero settlement calls. This adds the suggested one-time completion gate.

Changes

ga.py

• do_no_tool() final branch: at/above the threshold a normal completion now delegates to the existing do_start_long_term_update() (yield from, no duplicated prompt/SOP logic) and returns its settlement StepOutcome instead of next_prompt=None. The user's final answer is already streamed before tool dispatch, so nothing is hidden.
• _needs_long_term_settlement() + _in_autonomous_flow(): gate only fires when current_turn >= 15, settlement hasn't started, --no-user-tools is absent (same argv detection as agentmain.BANNED_TOOLS), and the latest [USER]: message isn't an [AUTO] autonomous prompt.
• do_start_long_term_update() sets _lt_started = True when it actually starts (after the < 10 refusal). This gives all three guarantees from the issue: an explicit prior call blocks the gate, a refused early call does not, and the settlement phase's own completion cannot re-trigger the gate (no recursion).

assets/tools_schema_cn.json

• 超15轮 → 达到15轮（含）, matching the EN schema (15+ turns) and the implemented turn >= 15 boundary — the issue's threshold-wording item.

Tests

frontends/tests/test_long_term_settlement_gate.py — 10 tests, one per case the issue lists: turn 14 exits normally; turn 15 / turn 40 enter settlement exactly once; settlement completion exits without recursion; a prior explicit call blocks the gate; a refused early call does not; no-op evaluation exits cleanly; --no-user-tools exempt; autonomous flow exempt (with latest-user-message detection). Tests use the real GenericAgentHandler — ga.py imports with stdlib only.

Full suite: 308 passed; the 3 failures (test_data_backup symlink ×2, test_release_qualification) are pre-existing Windows-env issues, identical before and after.

Scope notes

• Generic agent loop / agent_loop.py untouched — the gate lives entirely in GenericAgentHandler, as the issue suggested.
• Evaluation outcome is unchanged: the settlement prompt still allows a no-op conclusion; the gate forces evaluation, not memory modification.