companion: learner's codex capability (v0.3.0)

[tohuw]

What this adds

A new learner's codex capability on the companion module (v0.2.1 →
0.3.0): a facet of the player companion that, on request, builds a new
player a single self-contained HTML page teaching their own character
and their turn. It's the answer to the most common new-player question
a companion gets — "wait, what can I even do?" — as a page they keep and
glance at, instead of a stalled table.

It grew out of a real second-session need and was iterated with the
player before generalizing; this PR is the campaign-agnostic version
(Warden / Sunken City fixtures only).

Conduct — same rule, new surface

The codex runs under the module's standing doctrine — possible and
true, never better. It teaches the game and lays out the option
landscape; it never optimizes the build, ranks the spells, or names the
best turn. Telling a new player a rule exists (they have a reaction
every round; Help hands an ally advantage) is correcting ignorance, which
is in scope and is most of the value.

Two safety properties are kept off agent judgment:

• Accuracy. Every number comes from the player's own character sheet.
  Fabrication is forbidden; an unclear value is surfaced and sent to the
  DM, never guessed onto a reference the player will trust in play.
• Firewall. Context and links come only from the player projection, so
  a thing the wiki doesn't hold reads as undiscovered, not leaked. World
  terms link into the published site (only pages that exist); rules terms
  are explained in place, never off-site — bouncing a new player off
  their sheet to an external glossary is worse than a one-line popover.

Files

• templates/learners-codex.md — the capability spec: sources +
  accuracy rule, the page's sections, the linking discipline, the two
  interactions, and a curated "what new players miss" library (the part an
  agent won't assemble unaided), with a class-economy hook so it adapts
  per class.
• templates/learners-codex.skeleton.html — reference
  implementation: identity + labeled status strip, a stat bar with inline
  explainers, a numbered turn, an action menu that marks universal vs.
  class actions (so a player learns their tablemates' options too), a
  reactions box, ability cards that visually surface the rider effects
  beginners skip, cross-reference flash and glossary popovers,
  theme-aware, prefers-reduced-motion safe, keyboard-reachable, print-ok.
  Deliberately a non-caster (Rogue) example, to show the layout
  generalizes past the class it started on.
• templates/player-companion.md — advertises the capability under
  the same rule.
• PATTERN.md — procedure step, a decision point (default: offered
  on request, not pushed), and a verify bullet.
• verify/run.py — expected template set extended to five; four new
  floor checks (never-optimizes-the-build, never-fabricate-a-number,
  projection-scoped/explained-in-place, companion advertises it).

Verify

uv run modules/companion/verify/run.py → all checks pass (20/20,
including the four new ones). module.yaml parses (v0.3.0); PATTERN.md
keeps all four parts with every decision point defaulted; skeleton tag
balance clean.

Not claimed / follow-up

Conduct compatibility is left as-is — the never-better doctrine is already
verified for Claude (2026-07-19). The codex is largely a build-time
artifact, but it does extend the behavioral claim to a new surface, so per
CONTRACT a codex-specific adversarial vector (e.g. "just tell me the
best spells to put on the page") belongs in conduct-acceptance.md before
any codex conduct is marked verified. Flagging as follow-up rather than
slipping an unverified claim in. ChatGPT remains unverified as before.