investigate(vcon): log caller stack + coerce str arg in Vcon.__init__

[pavanputhra]

Why

A production conserver has been emitting a continuous stream of conserver.link.count{link_name=tag, outcome=error} with the same exception:

TypeError: string indices must be integers, not 'str'
  File "conserver/links/tag/__init__.py", in run
    vCon.add_tag(tag_name=tag, tag_value=tag)
  File "common/vcon.py", in add_tag
    tags_attachment = self.find_attachment_by_purpose("tags")
  File "common/vcon.py", in find_attachment_by_purpose
    for a in self.vcon_dict["attachments"]
             ~~~~~~~~~~~~~~^^^^^^^^^^^^^^^
TypeError: string indices must be integers, not 'str'

self.vcon_dict is a str at failure time. But:

• Every sampled vCon in Redis (JSON.TYPE) is object, with .attachments an array.
• Local repro with the same dict shape succeeds.
• The deployed code is current with fix(vcon): decode body per spec on read, canonicalize on write #182 and fix(links): use pydash.get + decode body to fix analyze chain (CON-573) #184.

So a caller is reaching Vcon.__init__ with a JSON-encoded string instead of the parsed dict. The current __init__ (self.vcon_dict = json.loads(json.dumps(vcon_dict))) silently round-trips a str, leaving self.vcon_dict as a str. The crash only surfaces on the first dict-style access — typically find_attachment_by_purpose from the tag link.

Static analysis didn't find the bad caller. This PR adds the instrumentation that will.

What this change does

1. Detect isinstance(vcon_dict, str) at the Vcon.__init__ boundary and log an ERROR with stack_info=True. The next failure to land will show the originating call site.
2. Coerce valid JSON-strings back to a dict so the chain doesn't crash and vCons don't pile up in the DLQ during the investigation. Non-JSON input falls back to {} rather than poisoning downstream access.

The defensive coercion is meant to live only as long as it takes to identify the caller and fix it at source.

Test plan

• pytest tests/core/test_vcon.py — 36/36 pass, including two new regression tests:
  • test_init_coerces_json_string_arg_and_logs_caller
  • test_init_bails_to_empty_dict_for_non_json_string
• Deploy and grep logs for Vcon.__init__ received a str lines + their stack — should identify the bad caller within minutes.
• Verify conserver.link.count{link_name=tag, outcome=error} rate drops to ~0 after deploy (defensive coercion takes over).

🤖 Generated with Claude Code