Adopt capsize_commons.logging in spikeforge (B″, hq#3)

[w4ffl35]

Closes Capsize-Games/hq#3 (B″). Deletes the duplicated observability/json_formatter.py and renders through capsize_commons.logging.JsonFormatter.

Changes

• Delete spikeforge/observability/json_formatter.py; logging_setup.py now builds JsonFormatter(message_key="event", identifier_fields=("run_id","config_id","config_hash")). LOGGER_NAME and the event key are unchanged.
• Clear the import blocker. Both spikeforge/__init__.py and spikeforge/observability/__init__.py resolve their eager imports lazily (PEP 562). The issue's blocker note named the chain (spikeforge/__init__.py → torch) and the observability eager imports; fixing only the latter would not have made the logging surface torch-free, so both were made lazy. Public names are preserved (verified below).
• Manifest: capsize-commons>=0.1.1 on the core distribution, plus the exact-pin update in test_packaging_profiles.py and documentation/project-layout.md.

Acceptance criteria

#	Criterion	Result
1	logging surface imports without the ML stack	✅ import spikeforge.observability.logging_setup — torch/snntorch not in sys.modules; all 7 __all__ names present
2	json_formatter.py deleted, callers use the library	✅ only caller (logging_setup.py) switched; message_key="event" + 3 identifier fields
3	pytest tests/test_observability_logging.py passes	✅ 5 passed
4	before/after dump with identical keys+values	✅ see below
5	capsize-commons declared; ruff+mypy clean	✅ ruff clean repo-wide; mypy 412 files, no issues
6	LOGGER_NAME still importable from ...logging_setup	✅

Behaviour-parity artefact

Same log record (train.started, run_id=run-1, config_hash=abc, fields={"epoch":2}), timestamp frozen:

before  {"timestamp": "2023-11-14T22:13:20+00:00", "level": "INFO", "event": "train.started", "logger": "spikeforge", "run_id": "run-1", "config_hash": "abc", "fields": {"epoch": 2}}
after   {"timestamp": "2023-11-14T22:13:20+00:00", "level": "INFO", "logger": "spikeforge", "event": "train.started", "run_id": "run-1", "config_hash": "abc", "fields": {"epoch": 2}}

json.loads(before) == json.loads(after) → True. The only difference is JSON key order (logger now precedes event); JSON objects are unordered, so this is not a semantic change.

Edge deltas found by direct comparison (reported, not hidden):

Case	before	after
fields={}	key omitted	"fields": {} emitted
exc_info set	no exception key	"exception": "<traceback>" added
non-ASCII message	\u00e9-escaped (ensure_ascii=True)	raw UTF-8 (ensure_ascii=False)

All are additive or parse-equivalent; none renames event or drops a correlation field.

Test results

• tests/test_observability_logging.py → 5 passed (torch-free venv, capsize-commons==0.1.1).
• Full suite in that env → 118 passed, 6 skipped, 5 failed, 154 errors. Every failure/error is environmental:
  • the 154 errors are ModuleNotFoundError: torch on collection (the suite requires the ML stack);
  • the 5 failures are tests/test_version_reporting.py (torch absent + spikeforge not pip-installed in this venv).
  • Attribution: git stash + re-run on the pre-change tree makes test_version_reporting.py fail harder (collection ERROR, not 5 assertion failures) — the change is not the cause.

Not verified

• The full suite was not run to green — it cannot run without torch in this tree (no .venv/uv.lock). CI owns that.
• snn_interpreter was not touched (provenance, hq#2, found no unique work to port).
• No publish/release performed.

[w4ffl35]

Compensating local run — recorded owner-approved bypass (hq#14)

Owner decision on Capsize-Games/hq#14: option 1 (path-filtering) + a recorded one-off bypass for this PR and airunnerweb#280. No red check is merged past — this PR is 0 failed; it is blocked only by the queued spikeforge-ci self-hosted pool (run 35459161595 queued since 17:47:21Z with no progress; the last successful run took 2h17m).

Compensating run, this worktree, capsize-commons==0.1.1:

Check	Result
pytest tests/test_observability_logging.py tests/test_packaging_profiles.py	31 passed, 2 skipped
ruff check .	All checks passed
mypy --ignore-missing-imports spikeforge/ … server/	Success: no issues found in 412 source files
import spikeforge.observability.logging_setup	torch-free; LOGGER_NAME=spikeforge

The full suite still requires the ML stack and is not run here (documented in the PR body).