Add SRD dungeon stocking and the KeyedEncounter.hoard gate (1.3.0)

[mmacy]

Grows osrlib with the SRD dungeon-stocking procedure its own spec deferred with "the generator that consumes them comes later" — osr-editor's phase 7 is that first consumer — and cuts 1.3.0.

Added

• osrlib.crawl.stocking.stock_area — the stocking procedure over the shipped tables. Given a dungeon level, an effective monster catalog, and one RngStream, it rolls one keyed area's contents and answers a frozen StockedArea of content models (a KeyedEncounter, an unguarded AreaTreasureSpec, or a StockedNpcParty report) an author can review, place, and edit. Every draw comes from the passed stream in a fixed order — contents d6, treasure d6 when the row's chance is non-zero, then on a monster room the d20 row, its count, and the variant roll or per-individual pool picks — so a stocked area is reproducible from the stream's state alone.
• KeyedEncounter.hoard (default True) — gates _generate_lair_hoard at first spawn. The default preserves every existing document's play semantics; hoard=False expresses the treasure-absent keyed room the SRD stocking roll produces, which an unconditional lair hoard could not otherwise represent.

Shared resolution, not a parallel copy

select_encounter_individuals is extracted into core/tables.py as the single home for resolving a monster row's individuals (variant / pool / single) from a stream. Both wandering_check and stock_area consume it, so a stocked row can never drift from a wandering roll on that row — the byte-for-byte agreement is structural rather than a hand-kept parallel. The duplicated _dice_minimum is gone with it.

Boundaries

Traps and specials produce no models (B/X ships example lists as referee prose, not tables), and an NPC-party row has no authorable content model — the procedure reports the rolled kind and count and stops. Treasure generates exactly once at play: a stocked monster room's hoard comes from the encounter's gated lair cache, an empty or trap room's from its unguarded spec, never both.

Tests

tests/test_stocking.py — seeded goldens across all four contents kinds, the variant / pool / NPC-party rows, count-pinning, the catalog-resolution contract, per-address stream independence, the shared resolver's draw, and the hoard gate driven through a real session (same seed, gate flipped: exactly one cache vs. none).

Full gate green: ruff format, ruff check, pyright 0 errors, 1579 passed / 58 skipped.

https://claude.ai/code/session_01SN8gc3PnSwXgKqqgfVcA4x