Review feedback (phase-rs#6274): the non-vacuity floors shipped at 2845/1480 sat
exactly at the PRE-refresh catalog, so reverting this PR's regen still
passed every assertion — the guard could not detect losing the refresh it
was added alongside. The pre-PR test asserted the parent counts as
equalities (assert_eq!(total_tokens, 2845)) and was green on main, so the
parent catalog measures exactly those floors.

Ratchet all floors to the committed vintage (2858/1490) and add the third
axis already carried on the summary, source_card_refs >= 9821. `>=` still
keeps weekly upstream ADDITIONS green — the false-red phase-rs#6237 introduced —
while any shrink now fails.

The ref floor is the load-bearing one and is deliberately tight: across the
ten recorded revisions of known-tokens.toml, tokens and rules_text are
monotone, and source_card_refs shrank exactly once — phase-rs#6199, a parser PR
that silently dropped 1173 token<->card links (9810 -> 8637) while the
token count GREW past a count-based floor. This regen is what repaired it.

Each floor is probed in isolation, since sequential assert!s let an earlier
conjunct dominate a later one and render it vacuous:
  - catalog reverted to f0ac543^   -> FAIL "token catalog shrank: 2845 presets < 2858"
  - 1 rules_text line removed        -> FAIL rules_text_tokens >= 1490
  - 5 source_card_refs blocks removed -> FAIL source_card_refs >= 9821
Catalog restored byte-identically after each probe.

Floors carry the measured count in the panic so a CI red names the axis and
the delta instead of requiring a local repro.

Assisted-by: ClaudeCode:claude-opus-4.8

Review feedback (phase-rs#6274): the non-vacuity floors shipped at 2845/1480 sat
exactly at the PRE-refresh catalog, so reverting this PR's regen still
passed every assertion — the guard could not detect losing the refresh it
was added alongside. The pre-PR test asserted the parent counts as
equalities (assert_eq!(total_tokens, 2845)) and was green on main, so the
parent catalog measures exactly those floors.

Ratchet all floors to the committed vintage (2858/1490) and add the third
axis already carried on the summary, source_card_refs >= 9821. `>=` still
keeps weekly upstream ADDITIONS green — the false-red phase-rs#6237 introduced —
while any shrink now fails.

The ref floor is the load-bearing one and is deliberately tight: across the
ten recorded revisions of known-tokens.toml, tokens and rules_text are
monotone, and source_card_refs shrank exactly once — phase-rs#6199, a parser PR
that silently dropped 1173 token<->card links (9810 -> 8637) while the
token count GREW past a count-based floor. This regen is what repaired it.

Each floor is probed in isolation, since sequential assert!s let an earlier
conjunct dominate a later one and render it vacuous:
  - catalog reverted to f0ac543^   -> FAIL "token catalog shrank: 2845 presets < 2858"
  - 1 rules_text line removed        -> FAIL rules_text_tokens >= 1490
  - 5 source_card_refs blocks removed -> FAIL source_card_refs >= 9821
Catalog restored byte-identically after each probe.

Floors carry the measured count in the panic so a CI red names the axis and
the delta instead of requiring a local repro.

Assisted-by: ClaudeCode:claude-opus-4.8

…hot vintage-proof (#6274)

* fix(engine): commit W30 MTGJSON token catalog and make coverage snapshot vintage-proof

The weekly MTGJSON refresh introduced by #6237 (vintage gate 2026-07-20)
unfroze a stale token catalog cache. Deterministic regen (byte-identical
across two checkouts) against the 2026-07-20 MTGJSON data adds 13 tokens
(12 Secret Lair Drop printings: Food x8, Treasure, Ooze, Elf Warrior x2;
1 HOB Goblin Army) and 1184 [[token.source_card_refs]] (8637 -> 9821),
and picks up MTGJSON's face-name change 'Undercity // The Initiative' ->
'Undercity'. No tokens removed; coverage stays complete (2858/2858
supported, 1490/1490 rules-text tokens parsed).

analyze_token_coverage_treats_source_defined_pt_as_represented pinned
vintage-dependent absolutes (2845/1480), which go red every week upstream
adds tokens even at 100% coverage. Replace them with invariants
(supported == total, parsed == rules_text) plus non-vacuity floors at
the last-known-good baseline so an empty or truncated catalog cannot
pass vacuously.

Assisted-by: ClaudeCode:claude-fable-5

* test(engine): ratchet token-catalog floors to the committed vintage

Review feedback (#6274): the non-vacuity floors shipped at 2845/1480 sat
exactly at the PRE-refresh catalog, so reverting this PR's regen still
passed every assertion — the guard could not detect losing the refresh it
was added alongside. The pre-PR test asserted the parent counts as
equalities (assert_eq!(total_tokens, 2845)) and was green on main, so the
parent catalog measures exactly those floors.

Ratchet all floors to the committed vintage (2858/1490) and add the third
axis already carried on the summary, source_card_refs >= 9821. `>=` still
keeps weekly upstream ADDITIONS green — the false-red #6237 introduced —
while any shrink now fails.

The ref floor is the load-bearing one and is deliberately tight: across the
ten recorded revisions of known-tokens.toml, tokens and rules_text are
monotone, and source_card_refs shrank exactly once — #6199, a parser PR
that silently dropped 1173 token<->card links (9810 -> 8637) while the
token count GREW past a count-based floor. This regen is what repaired it.

Each floor is probed in isolation, since sequential assert!s let an earlier
conjunct dominate a later one and render it vacuous:
  - catalog reverted to f0ac543^   -> FAIL "token catalog shrank: 2845 presets < 2858"
  - 1 rules_text line removed        -> FAIL rules_text_tokens >= 1490
  - 5 source_card_refs blocks removed -> FAIL source_card_refs >= 9821
Catalog restored byte-identically after each probe.

Floors carry the measured count in the panic so a CI red names the axis and
the delta instead of requiring a local repro.

Assisted-by: ClaudeCode:claude-opus-4.8

* docs(engine): document the provenance seam for the token-catalog floors

The floors observe totals; provenance (committed catalog == tokens-gen output
for the declared MTGJSON vintage) is structurally untestable at test time (the
~560 MB gitignored generator input is absent; the test sees build.rs's embed of
the tracked file) and is verified at the generation seam instead
(gen-card-data.sh temp+cmp+vintage gate). Comment documents the seam and the
by-hand reproduction recipe, per review 4746023659.

Assisted-by: ClaudeCode:claude-opus-4.8