fix(timeline,tools): GPT-5 timeline robustness — reasoning-item drop + tool-batch isolation

[ngoclam9415]

Two independent robustness fixes for GPT-5/o3/o4 timelines. Both surfaced while verifying a real session timeline (atlas-q1).

---

Commit 1 — persist reasoning items on empty-summary turns

Problem

GPT-5/o3/o4 turns return a reasoning item (rs_… + encrypted_content) with an empty summary on low-summary turns — typically single-call tool continuations. _record_think_results gated AGENT_THOUGHTS creation on summary text:

if reasoning and len(reasoning) > 0:   # keys on summary text

When the summary was empty, no entry was created — and reasoning_items (which rides in that entry's metadata) was silently dropped along with it. Observed: 7 of ~14 model turns in atlas-q1 were bare tool_call entries with no preceding agent_thoughts.

Impact: on resume, those turns replay with no reasoning state. Cross-turn reasoning replay (PR #10) silently does not apply to low-summary turns.

Fix

Gate now keys on reasoning_items presence, not summary text. Entry emitted with empty content when only the item is present; metadata still carries the item for replay. Applied to both branches.

---

Commit 2 — isolate tool-call failures within a batch

Problem

A dispatch-phase exception (registry getattr, name parsing, object lookup) escaped before the inner try in _execute_single_call / _execute_single_call_async. In the async path it propagated out of asyncio.gather, discarding the entire batch's results — including calls that succeeded.

The TOOL_CALL entry already recorded N tool_call_ids, so the next OpenAI turn 400s on the unanswered tool calls.

Fix

• Each single-call dispatcher wrapped in one outer guard covering dispatch + execution; both now non-raising. Redundant inner try blocks removed.
• execute_tools_async: asyncio.gather(return_exceptions=True) + escaped-exception → error result (defense-in-depth).

Every tool_call_id in a batch now always gets a result.

---

Tests

• TestEmptySummaryReasoningPersistence — tool-call / direct-answer branches + no-items negative case.
• Batch-isolation tests — async + sync paths: failing call isolated, siblings succeed, tool_call_ids preserved.
• Followed RED→GREEN for both.
• test_thinking_metadata_persistence.py 14/14, test_tool_executor_parallel.py 9/9, tests/unit/core/ + tests/regression/ 202 passed / 18 skipped.
• 2 pre-existing failures in TestEndpointHashAndFingerprint (test pollution, fail on clean tree too) — out of scope.

Notes / not in scope

• Mixed async/sync tool batches still degrade (sync tool blocks the event loop). That's a separate design item — needs tool concurrency classification — deliberately not bundled here.
• Existing broken timelines (e.g. atlas-q1) stay unrecoverable; commit 1 fixes capture going forward only.