fix: three papercuts the end-to-end run turned up

[vicenteliu]

Three smaller defects from the same end-to-end run. Unrelated to each other, all small, all real.

1. Editing the model list from the admin UI deleted the comments inside it

Found by using the product as intended: adding two models through PUT /api/admin/playbook-models stripped three comments from the checked-in playbook, including

    params:
      # Sonnet 5 / Opus 5 reject temperature / top_p / top_k (HTTP 400).
      max_tokens: 8192

which exists precisely to stop someone re-adding a field that 400s.

write_playbook_models splices the file around the model: and extra_models: blocks, so comments outside survived — but both blocks are re-rendered from parsed data, so comments inside could not. The function docstring claimed flatly "All other content — comments, key order, formatting — is preserved."

Two ruamel details made this harder than it sounds, and I got the first fix wrong before catching it:

• A comment before the first key of a nested mapping attaches to the parent's key. The one above max_tokens: is stored on params, not on max_tokens.
• The heading of the next top-level section is parked on the last key of this block. Carrying it forward printed # Loop bounds. twice — once where it belongs and once inside the rewritten block. Column separates them cleanly: an inner comment is indented (col=6), a section heading sits at col=0.

An unchanged value now also keeps the scalar ruamel parsed, so an edit no longer restyles effort: "high" into effort: high. An unchanged rewrite is byte-identical, which is now a test.

2. --model on workingset distil could not switch providers

model_name = model or target.name
provider = make_provider(target.provider_id, kind=target.kind)   # ← unaffected by --model

It replaced only the model name and kept the tier's provider. So passing the id that GET /api/models returns and the UI uses:

$ opspilot workingset distil ws_… --model anthropic/claude-sonnet-5
ProviderError: Anthropic API error: 404 - {'type': 'not_found_error',
  'message': 'model: anthropic/claude-sonnet-5'}

— as a full Python traceback, because only NotDistillableError was caught. And naming another provider's model would have sent it to the wrong API entirely.

It now resolves against the playbook's model list, accepting either provider/name or the bare name, refuses an unknown one with the list of valid ids, and exits cleanly on ProviderError. Both verified against the live instance.

3. The API dropped working_set_id from every Consultation

_con() serialized id / author / title / created_at / updated_at / pinned_reason / session_id — but not working_set_id.

ADR-0036 makes the Working set the chain of Consultations on one problem, and that chain is what distillation reads, so this was the domain's central relationship missing from its own API. Nothing was broken at the time — distillation calls consultations.for_working_set() on the store directly and the web client never declared the field — which is exactly why it survived.

Verification

6 new tests. The playbook ones use a fixture with comments inside the model blocks; the existing fixture put every comment outside them, which is why it stayed green while the real file lost comments on every edit.

pytest -m "not slow and not requires_ollama and not requires_api_key" — 1317 passed. No protected path touched.

🤖 Generated with Claude Code