v1.29.1

No rule changes: a sweep review's findings applied as text. Contradictions between
sections reconciled, claims trimmed to what the scripts actually do, stale counts and
dead cross-references corrected, and the skill body cut back below the validator's
ceiling.

Changed

• Contradictions: three places said the knowledgebase is read "alongside" or "not ahead
  of" the engine's projection while section 10's rule has the read done first and the
  case and the number said together; all three now say "said beside the engine's
  projection, never instead of it". Section 5's settings-blind rule now carries the
  exception section 4 already states (a host tool the settings panel shows scored to
  this league). Section 6's post-draft IR step now runs the stash procedure at once and
  again on the clear date, which is what section 8 and the reference say.
• Claims the scripts do not keep: the refresh seed is described as seeding the role
  collector from the starters whose absence plan is still open, not "the collectors"
  from every - cell (skill section 10, and the sentence removed from
  scripts/kb-lint.py's docstring); README names the nine rollup files rather than
  "every NFL-wide file"; scripts/pull-list.py --check no longer claims to check dates;
  the knowledgebase reference names the rollup outputs the script actually writes.
• scripts/merge-tags.py reads data/teams.md the way the other three readers do, a
  leading and trailing pipe stripped, so a markdown-table row is accepted.
• Stale text: the build's agent count, the burn-the-claim test's contention input (the
  host's transaction trends, per section 4), the joins list, and the location the
  scripts are copied to and run from (the knowledgebase's build/ folder, in the
  reference's every command and in README).
• Dead cross-references: the draft-time IR value test points at itself and section 7's
  dart forecast rather than a value test section 7 does not hold; scripts/kb-lint.py
  and scripts/build-scaffold.py cite the templates in references/knowledgebase.md
  instead of a build/templates-lean.md the plugin does not ship; the refresh's
  runnable form is described as the user's own workflow file; the validator is named;
  and the draft-time comparator names the healthy player the pick would otherwise take,
  which is what "the bench player he displaces" meant at a draft.
• Bloat, every rule kept: the scoring editor's field inventory moves verbatim from
  SKILL.md section 6 to references/draft-sharks.md under its own heading, with the
  audit rules and the pointer left in the skill; section 10's prose copy of the
  knowledgebase layout becomes one clause per set; the IR bullet drops the restatement
  of the section 10 read it already points at; and the reference's guide-pass step, its
  build history clauses and README's fill-check paragraph defer to the scripts'
  --help.

Why: a sweep review of 1.29.0 read every rule against every other rule and against the
shipped scripts. Nothing it found changed a rule; what it found was text that had drifted
from the rules and from the code, and a skill body 489 lines into a 500-line ceiling with
no room for the next rule.