Add RollDice authorial adjudication command + release 1.1.0 (#22)

[mmacy]

Closes #22.

Adds an authorial RollDice command so a referee front end (the osrlib-referee MCP server) can resolve a freeform chance outcome by rolling through the seeded session — accepted, logged, replayable, and grounded in a typed event — instead of inventing a number or rolling out-of-band. This completes the authorial surface for freeform adjudication described in docs/front-ends/llm-referees.md.

What's in scope

RollDice only. Per the plan's Correction 1, Check and Save are deferred to their own follow-up issues (Save is not a thin wrapper — combat.saving_throw is entity-coupled). RollDice alone unblocks the front end.

Design

• Command — RollDice(expression: str) in crawl/commands.py, appended to ALL_COMMAND_CLASSES (stable, additive wire order) and reachable via AnyCommand / parse_command. Legal in every mode (inherits _ALL_MODES).
• Validation surface — Option A (the plan's recommended default). A malformed expression is rejected at construction by a @field_validator calling dice.parse, exactly as SpawnMonsters.count_dice is. This keeps Rejection reserved for state/legality and matches the existing dice-string precedent. This deliberately supersedes the issue's literal acceptance-criterion wording ("returns accepted=False with a Rejection"): a bad expression surfaces as a parse-time ContentValidationError the front end catches, not a CommandResult. The RNG draw happens strictly after validation, so a malformed roll consumes no draw and no log entry — the same rejection purity, achieved one layer earlier.
• Dedicated stream — a new ADJUDICATION_STREAM = "adjudication" keeps ad-hoc referee rolls off the mechanical streams, so an adjudication roll never perturbs the draw sequence of a keyed mechanic (encounter/combat/exploration) and keyed outcomes stay fixed under replay.
• Event — DiceRolledEvent (referee visibility) carries expression, total, and the individual rolls, with a stable adjudication.dice_rolled message code and a default format_message template.

Deliberate deviation from the plan's touch-point table

The plan's table listed defining DiceRolledEvent in core/events.py but registering it in crawl/events.py's CRAWL_EVENT_CLASSES. I instead define and register it entirely in crawl/events.py. Every other CRAWL_EVENT_CLASSES member emitted by a referee handler (FlagSetEvent, MonstersSpawnedEvent, TimeAdvancedEvent, GameOverEvent) is defined in crawl/events.py; DiceRolledEvent is emitted only by the crawl session handler _handle_roll_dice, so a split home (define in core, register in the crawl list) would make it the lone crawl-registered event living in core — a "one home per symbol" violation (AGENTS.md greenfield discipline). No osrlib.core function needs to emit it, so core placement buys nothing and the layering rule is upheld either way.

Tests

New coverage maps to each acceptance criterion: accept + logged (TestRollDice), malformed-expression rejection with no draw / no log, referee visibility, byte-identical determinism under a fixed seed, stream isolation (both a "touches only the adjudication stream" diff test and an interleaved-roll-leaves-SpawnMonsters-unshifted test), save/load + command-log replay (test_persistence.py), and legality in every mode. Fixture wiring added to test_commands.py, test_events_kernel.py, and the property-based command fuzzer (test_crawl_properties.py). Docs (rng-streams.md) and CHANGELOG.md updated.

ruff format --check, ruff check, pytest (1512 passed, 58 skipped), and a strict mkdocs build are all clean.

Review provenance

Followed the AGENTS.md plan → rubber-duck → PR loop; the issue body + owner comment are the ducked plan. A fresh skeptical reviewer audited the implementation against AGENTS.md invariants (determinism, layering, command purity, events/schema discipline, greenfield anti-patterns) and the acceptance criteria. Verdict: SOLID, no blocking findings. It endorsed the event-home deviation as the more consistent choice, confirmed the stream-isolation tests are non-tautological, and verified the message template renders honestly when total != sum(rolls). Its two non-blocking notes (name the AC-wording supersession explicitly; record the event-home deviation) are folded into this description.

[mmacy]

Folded the 1.1.0 release into this PR (following the v1.0.0 precedent of bumping within the impl PR and tagging its merge commit):

• pyproject.toml → 1.1.0 (minor: additive public command/event surface), uv.lock regenerated.
• CHANGELOG [Unreleased] → [1.1.0] - 2026-07-05 with refreshed compare links.
• Local release dry run passed: uv build, tools/release/check_dist.py dist 1.1.0, and a fresh-venv install_smoke.py 1.1.0 all clean.

Versioned-docs (mike) adoption — the AGENTS.md trigger for this first post-1.0 minor with API changes — is consciously deferred to keep the release focused; tracked in #24.

On merge, an annotated v1.1.0 tag on the merge commit fires release.yml (gate → build → dual-OS audit/smoke → PyPI trusted publish → GitHub Release).