Resolve B2B targeting facets into LinkedIn targetingCriteria URNs

[th0rz05]

Summary

• Adds TargetingResolver, which turns the LLM-extracted free-text B2B facet names from a brief (seniorities, job functions, industries, titles, skills, company sizes) into the entity URNs LinkedIn's targetingCriteria requires.
• Picks a resolution path per facet based on what the API actually exposes:
  • Closed enums (seniorities, jobFunctions) — no typeahead finder; resolved against the standardized /seniorities and /functions lists by display name. Lists are fetched once and cached per publish.
  • Open taxonomies (industries, titles, skills) — resolved via the typeahead finder, preferring an exact normalized name match, else LinkedIn's top-ranked hit.
  • company_sizes — mapped statically to urn:li:staffCountRange:(min,max) tuple URNs (not the SIZE_* enum the facet listing returns; the top bucket uses INT_MAX as its upper bound).
  • locations — from the static ISO→geo map.
• No-guess contract: any name that resolves to no real LinkedIn URN is surfaced in ResolvedTargeting.unresolved (caller flags it as a manual Campaign Manager step) and is never fabricated into a URN.
• Wires the resolver into the publish flow (server.py), replacing the static geo-only targeting. mapping.audience_to_targeting is trimmed to the geo-only baseline it now is, retained for inspection/tests.
• Adds three read-only client methods: typeahead_targeting_entities, list_seniorities, list_functions.

Verification

• 41 unit tests pass; ruff clean on touched files.
• Verified live against a DRAFT campaign: the full resolved targetingCriteria (staffCountRange tuple URNs + seniorities, jobFunctions, industries, skills, locations) was accepted by LinkedIn and read back from the API. All test resources and a prior-session orphan were deleted; account confirmed clean. Nothing can spend — everything stayed DRAFT.

Test plan

• pytest -q green
• ruff check clean
• Live DRAFT publish accepted by LinkedIn + torn down

[Copilot]

Pull request overview

Adds end-to-end resolution of free-text B2B audience facets into LinkedIn targetingCriteria URNs, and wires that resolver into the LinkedIn publish flow while preserving the “no-guess” contract (unmatched values are surfaced, not fabricated).

Changes:

• Introduces TargetingResolver to resolve seniorities/job functions via standardized lists and industries/titles/skills via typeahead, plus static staffCountRange and ISO→geo mapping.
• Updates LinkedIn publish flow to use resolved targetingCriteria and report unresolved facet values as publish notes.
• Adds LinkedIn client helpers for typeahead + standardized list endpoints and a focused integration test suite for resolver behavior.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
tests/integrations/test_linkedin_targeting.py	Adds resolver integration tests for enum vs typeahead paths, no-guess behavior, staffCountRange mapping, and caching.
src/yieldagent/integrations/linkedin/targeting.py	New resolver module that converts Audience facets into LinkedIn targeting URNs with unresolved surfacing.
src/yieldagent/integrations/linkedin/server.py	Switches publish flow to use TargetingResolver output and updates unresolved targeting notes.
src/yieldagent/integrations/linkedin/mapping.py	Trims audience_to_targeting down to the geo-only baseline and updates its docstring accordingly.
src/yieldagent/integrations/linkedin/client.py	Adds read-only client methods for typeahead targeting entities and standardized seniority/function taxonomies.
briefs/linkedin_live_test_brief.md	Adds a live-test campaign brief used for manual verification/reference.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.