fix(backend/copilot): close empty-completion gaps + add silence heartbeat (#13056)

## Why

Three Discord reports in `#While testing the new payment, billing`
describe AutoPilot turns that "just stop":

- **John (May 6, dev)** session `07da6f3e-a183-4315-8b5e-a1c99f3b3d29`
was root-caused as SECRT-2333 and patched in #13052 (orphan
`browser_act` tool_use blocks on Kimi K2.6 + SSE drop).
- **Toran (May 6, prod)** screenshot shows an expanded `Reasoning ^`
panel inline with no answer text, no fold. SessionId not shared yet, but
the FE shape matches a turn that finished thinking-only without ever
emitting text.
- **Ubbe** intermittent — sometimes a refresh reveals "Credits limit
reached".

Investigation surfaced gaps that #13052 explicitly does not close:

- **SDK**: `_should_surface_empty_completion` covers ghost-success and
orphan-tool cases, but not `has_started_reasoning AND not
has_started_text AND not _any_tool_results_seen AND subtype="success"` —
the genuine thinking-only-no-tools path. Falls through to plain
`StreamFinish`, and the FE's `splitReasoningAndResponse` then renders
the reasoning inline (no fold), visually identical to in-progress
streaming.
- **Baseline path** (`baseline/service.py`): only emits silent-finish
fallback when `iterations >= agent_max_turns` (line 1996-2012). Natural
finish with empty terminal text has no equivalent.
- **Silence between events**: `StreamStatus` is emitted at boundaries
("Contacting the model…", "Analyzing result…") but never *during* long
inter-event silence. The FE rotates through generic phrases for up to
several minutes before the next chunk lands.

## What

Single PR with three tightly coupled changes:

1. **SDK thinking-only-no-tools guard** — extend
`_should_surface_empty_completion` in `sdk/response_adapter.py` to fire
when the model emitted only `ThinkingBlock` content and ended
`subtype="success"` with no tools. Surfaces as the same
`StreamError(code="empty_completion") + StreamFinish` pair as the
existing two cases.

2. **Baseline empty-finish fallback** — add
`_NATURAL_FINISH_EMPTY_FALLBACK_TEXT` +
`_natural_finish_empty_notice_text` +
`_build_natural_finish_empty_fallback_events` mirroring the
budget-exhausted helpers, plus a new branch in the streaming loop
covering `loop_result.finished_naturally AND terminal_round_text == ""`.

3. **Silence watchdog** — new `backend/copilot/stream_heartbeat.py` with
`SilenceWatchdog` async context manager + `wrap_stream_with_heartbeat`
generator. Integrated in `executor/processor.py` between the raw driver
stream and `stream_registry.stream_and_publish`, so both SDK and
baseline paths get the heartbeat. Default schedule:

| Threshold | Message |
|-----------|---------|
| 10s | Working on it… |
| 30s | Still working — complex requests can take a moment… |
| 60s | This is taking longer than usual… |

The watchdog re-arms after every non-status event and self-suppresses
for 5s after any driver-emitted `StreamStatus` so it doesn't talk over
messages like "Optimizing conversation context…".

## How

### Files changed

- `backend/copilot/sdk/response_adapter.py` — third branch in
`_should_surface_empty_completion`.
- `backend/copilot/sdk/response_adapter_test.py` — 2 new tests (positive
+ regression).
- `backend/copilot/baseline/service.py` — new constant, helpers, and
silent-finish branch.
- `backend/copilot/baseline/service_unit_test.py` — 6 new tests.
- `backend/copilot/stream_heartbeat.py` (new) — `SilenceWatchdog`,
`wrap_stream_with_heartbeat`.
- `backend/copilot/stream_heartbeat_test.py` (new) — 10 tests covering
threshold firing, bump reset, escalation order, one-per-gap dedupe,
suppression, cleanup, error tolerance, wrapper passthrough, wrapper
interleaving, wrapper suppression.
- `backend/copilot/executor/processor.py` — wires the watchdog around
the driver stream.

### Out of scope

- Underlying `browser_act` "tool never executed" bug — separate
tool-execution layer concern, per #13052 PR description.
- Auto-retry on `code="empty_completion"` — flagged in #5d1cdc2bad as
future follow-up.
- Mid-stream credit-balance check / surfacing — orthogonal billing
concern.
- FE rendering tweak so finalized-no-text-after-reasoning visually
distinct from streaming — depends on UX agreement (cancel button, retry
CTA, error card).
- Per-tool progress events (Approach C from the design discussion) —
heavier change, deferred.

## Test plan

```
poetry run pytest backend/copilot/sdk/response_adapter_test.py     # 66 pass (incl. 2 new)
poetry run pytest backend/copilot/baseline/service_unit_test.py::TestNaturalFinishEmptyNoticeText \
                  backend/copilot/baseline/service_unit_test.py::TestBuildNaturalFinishEmptyFallbackEvents \
                  backend/copilot/baseline/service_unit_test.py::TestBudgetExhaustedNoticeText \
                  backend/copilot/baseline/service_unit_test.py::TestBuildBudgetExhaustedFallbackEvents
                                                                   # 14 pass (incl. 6 new)
poetry run pytest backend/copilot/stream_heartbeat_test.py         # 10 pass (new)
poetry run pytest backend/copilot/sdk/service_test.py::TestStreamEndedWithoutResultMessage
                                                                   # 3 pass (regression)
poetry run black --check + poetry run ruff check                   # clean
```

- [ ] Manual QA on dev: trigger a thinking-only turn (extended_thinking
model, simple yes/no question that the model answers in thinking) and
confirm the FE now shows an error card instead of a stuck reasoning
panel
- [ ] Manual QA on dev: trigger a long tool call (e.g. video gen) and
confirm escalating heartbeat messages appear at ~10s/30s/60s
- [ ] Verify Toran's session if/when sessionId is shared — does this PR
cover his shape too?

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: majdyz <zamil.majdy@agpt.co>