feat(decks): rank owned-card upgrades by deck fit, not just legality (#294)

[untraceablez]

Closes #294. One of the two remaining items in the Decks & EDH tooling milestone; #295 is the sibling and reuses this scoring.

The problem

The heuristic suggester returned cards that were owned, in colour identity and Commander-legal, ranked by curve then price. Those rules make a pick allowed, not good — a generic in-colour mana rock could outrank a card that actually does what the deck is trying to do.

What changes

Candidates are now scored against the deck's plan first, with curve and price kept as tie-breakers. A deck with no detectable theme ranks exactly as it did before — only decks that have a plan see any difference.

Two signals, both offline and deterministic (no LLM, no network, no new data source):

• Themes — recurring keywords plus rules-text patterns (+1/+1 counters, tokens, sacrifice, lifegain, spellslinger, mill, graveyard recursion).
• Tribes — creature types the deck cares about, derived from the deck itself rather than from a hardcoded list: types the commander's rules text names, or that enough of the deck's creatures already share. Magic has hundreds of creature types and gains more every set, so a literal list would be wrong within a year; this stays correct for types invented after it was written. A tribal hit outweighs a text match, because it's the deck's whole plan rather than one mode of a card.

Whatever matched is shown on the pick — "matches your tokens, elf". An unexplained recommendation is one the reader has no way to disagree with, which is really the complaint behind this issue.

Shared plumbing, not a second copy

Theme detection moved out of llm.py into a new src/deck_synergy.py. It had to be shared: the deliberately-offline suggester can't import the LLM module to reuse it, and a second copy would drift. This repo has already paid for that once — two card-name resolvers disagreed until DFCs matched from a decklist but not from a CSV (#338). #295 asks for the same plumbing as its pre-filter, so it's there waiting.

Two things found while building it

• The plural of Elf is Elves. My first pass matched elfs?, which silently found nothing for Elf, Wolf, Dwarf and Ally — precisely the tribes people build decks around. Now only the irregulars real creature types actually use are handled, and an unknown one degrades to a missed signal rather than a wrong one.
• Subtypes are read from creature type lines only, so an Aura-heavy deck doesn't come out as "Aura tribal".

Deliberately not included

#294 lists a game_changer "staple-ness" boost as optional. I left it out on purpose: those cards raise a deck's bracket (src/brackets.py), so silently floating them to the top of an upgrades list would change something about the deck that the suggester never mentions. That's worth doing as an explicit, labelled signal — "strong, but raises your bracket" — rather than a hidden ranking nudge, and it's a decision worth making on its own rather than smuggling in here.

Verification

• 1274 passed, 100% coverage, 0 missed lines (16 new tests). ruff and mkdocs --strict clean. No migration.
• The behaviour tests are the point: a synergistic Ramp card outranks a generic one; synergy beats a cheaper, lower-curve generic; a themeless deck's ordering is unchanged; sideboard cards don't shape the profile.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LxrePqFWyzQTcWaroSk4sH