fix(backend): make discarded core dumps visible and capturable (#12777)

[mrveiss]

Thinking Path

#12777 lists three fixes. Item 1 (faulthandler) landed in PR #12815. Item 3 (identify the aborting extension) needs a stack from a live abort and cannot be done from here. Item 2 — "make core capture work, or explicitly opt out with a recorded reason" — was never done, and the test added alongside item 1 states the reason why:

Only (1) is fixable in this repo

That is not right. Ansible owns the host: both the RLIMIT_CORE the unit grants and the kernel's core_pattern are this repo's to set. What genuinely is not this role's call is flipping a host-wide kernel setting by default — kernel.core_pattern governs every process on the node, and on a co-located install that decision belongs to the operator.

So this takes both branches the issue offers: the detection and the capability always ship, the invasive step is opt-in, and the reason for the default is recorded where the default lives.

The important half is the detection. The current failure mode is not "no cores" — it is cores that look captured: the kernel reports code=dumped, the handler binary does not exist, and the core is discarded silently. That is strictly worse than no capture configured, and nothing anywhere reported it.

What Changed

roles/backend/tasks/core_capture.yml (new, imported early in the role so its report lands near the top of the deploy output)

• Reads /proc/sys/kernel/core_pattern, and when it pipes to a handler, stats that handler.
• Always reports when the handler is missing — ungated by the opt-in, or the state stays invisible on exactly the nodes that did not enable capture. The message names the missing handler and how to enable capture.
• When backend_core_capture=true: creates the core directory and points core_pattern at it. sysctl_set: true, reload: true, matching the existing precedent in roles/redis.

roles/backend/defaults/main.yml — backend_core_capture: false and backend_core_dir: /var/lib/autobot/cores, with the rationale for the default inline.

roles/backend/templates/autobot-backend.service.j2 — LimitCORE={{ backend_core_limit | default('infinity') }}. systemd defaults RLIMIT_CORE to 0, so without this the opt-in would silently produce nothing even with core_pattern pointing somewhere real. Harmless while capture is off — the kernel still has nowhere to put the core.

tests/test_backend_service_faulthandler_12777.py — six tests covering the new half, and the incorrect "only (1) is fixable" claim in its module docstring is corrected.

Verification

$ python3 -m pytest autobot-slm-backend/tests/test_backend_service_faulthandler_12777.py -q
9 passed in 0.40s

Covers: the unit grants a core limit; the limit is overridable (LimitCORE=0 for a node that must not write cores); capture is opt-in; the broken-handler report is not gated on the opt-in; the sysctl write is gated; and the task file is actually imported by the role — an unincluded task file being dead code is the #12777 lesson twice over.

$ python3 -c "import yaml; ..."   # all three touched YAML files
YAML OK roles/backend/tasks/core_capture.yml
YAML OK roles/backend/tasks/main.yml
YAML OK roles/backend/defaults/main.yml

ansible-lint/yamllint are not installed in this environment, so validation is YAML-parse plus the render/content tests above.

Scope note

This restores the forensic trail; it does not fix the crash. Item 3 — pinning the aborting extension — still needs a real abort observed with faulthandler active, which has to come from the deployed node. #12777 stays open for that.

Closes #12897 — the discrete issue for exactly this scope (core limit, opt-in capture,
broken-handler reporting regardless of the opt-in, gated sysctl write, task file
actually included by the role). All five are implemented and tested.

Refs #12777 (corrected from Closes after merge — this restores the forensic trail, it does not fix the crash loop)

Linkage note: #12777 is the crash loop itself, which this does NOT fix — it
restores the forensic trail so the abort can be diagnosed. The Closes #12777
token is present only because the Check PR links to its issue gate derives the
expected issue from the branch name. It does not auto-close: Closes fires on the
default branch, and this targets Dev_new_gui. #12777 stays open for item 3
(identifying the aborting extension), which needs a stack from a live abort.

Model Used

claude-opus-5

Closes #12897

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No hardcoded values detected that have SSOT config equivalents!