runneth-classic: infer + confirm naming conventions during onboarding

[rezakhadjavi]

What this PR adds

Naming convention inference + confirmation as part of runneth-classic setup.

Performance creative work lives or dies on naming conventions. Customers say "the May ads" or "the lip balm launch" or "the question hooks from the UGC test" — and the agent needs to translate those phrases into exact campaignName / adsetName / adName filters. Without a captured convention, the orchestrator either guesses (sometimes wrong) or runs motion meta filter-reference round-trips on every turn that contains a fuzzy phrase. Neither is great.

This PR makes naming convention discovery part of the standard onboarding flow.

Phase 0 addition

During the 365-day creative library pull (the same one that powers graveyard inference), Phase 0 also extracts structural patterns in:

• campaignName
• adsetName
• adName

It looks for common separators (_, -, |, /), consistent positional segments, repeated prefixes, and abbreviations. The goal is to decode each segment: market code, objective, product line, audience, hook tactic, format, iteration number, date code, persona, anything else encoded in names.

Saved to /agent/brain/runneth-classic/workspaces/<slug>/naming-conventions.md so the orchestrator's Step 0 re-anchor reads it on every future chain run.

Phase 1 addition — Turn 4d

Surface the inferred decoding as questions, using real examples from the workspace's actual data:

"Looking at your campaigns and ads, I think I'm seeing a pattern. Here's my read:

• Campaign names look like <MARKET>_<OBJECTIVE>_<PRODUCT>_<DATE> — e.g., US_Conv_LipBalm_0526 means US market, conversion campaign, lip balm product, May 2026.
• Ad set names seem to encode audience: <INTEREST>_<AGE>_<GENDER> — e.g., Wellness_25-34_F is the wellness-interest ad set targeting women 25-34.
• Ad names look like <HOOK_TACTIC>_<FORMAT>_<ITERATION> — e.g., Question_UGC_v3 is a question-hook UGC creative, third iteration.

Am I reading this right? Anything I'm missing — segments that mean something I haven't decoded, or codes that aren't obvious from the names alone?"

Customer corrects what we got wrong and fills in segments that data alone couldn't decode (e.g., "P1/P2/P3 = persona 1/2/3 in our strategy matrix").

The turn only fires if Phase 0 surfaced confident structural patterns. If names look random or human-readable, skip silently and note the absence in naming-conventions.md so the orchestrator knows to fall back to motion meta filter-reference for fuzzy phrases rather than guessing at structure.

How the orchestrator uses it

• Step 0 re-anchor reads naming-conventions.md on every chain run.
• Slice-integrity user.md rule references it when interpreting user filter phrases:
  • "the May ads" → campaignName includes "0526"
  • "the lip balm launch" → campaignName includes "LipBalm"
  • "the question hooks from the UGC test" → adName includes "Question" AND adName includes "UGC"
• Filter construction skips motion meta filter-reference round-trips when the convention decodes the user's phrase directly. Filter-reference becomes a fallback for cases the convention doesn't cover, not a default path.

Files changed

Only runneth-classic/setup-runneth-classic/SKILL.md. Adds 53 lines covering the Phase 0 extraction step and the new Turn 4d.

Cumulative iteration progress

This is the third PR layering improvements onto the original runneth-classic install after real test installs:

• Add follow-up setup skill fixes (Phase 0, graveyard, narrative close, one-question-per-turn) #94 (merged): Phase 0 pre-discovery, conversational one-question-per-turn flow, narrative close, integrations question reframed around creative-strategy outcomes, graveyard inference, current pressure, install_notes fetch guidance for large skill files
• runneth-classic: explicit brain handling policy (inventory, conflict resolution, archival) #96 (merged): explicit brain handling policy — inventory, conflict resolution, archival, provenance markers, Turn 1 branching by inventory state
• add-roles-permissions: v3.0.0 — six primitives, no mode toggle #99 (this one): naming convention inference + confirmation, so slice filters work without runtime round-trips

All against the same runneth-classic/setup-runneth-classic/SKILL.md. Each builds on the last.