docs(epistemic): say COHERENT 'claims/premises', not 'facts'

COHERENT is 'currently incident-free' and revisable (ADR-0021 non-hardening
invariant), not a hardened fact. Calling promotion premises 'facts' leaks an
implication that COHERENT = unrevisable truth. Replace:
- paper §3.4: 'reviewed facts' -> 'reviewed claims'; 'facts already marked
  COHERENT' -> 'claims already marked COHERENT'.
- issue: 'already-COHERENT facts' -> 'already-COHERENT premise set';
  'unreviewed facts' -> 'unreviewed claims'.

Doctrine unchanged. 'determined facts' (idle-tick term) left intact.
Green: epistemic invariants + phase3 + INV-21 (15 passed).

docs/issues/proof-carrying-coherence-promotion.md

@@ -20,10 +20,10 @@ the same change that raised this issue:
 1. **Doctrine precision (paper §3.4).** "Promotion requires a *curator-mediated*
    coherence judgment" over-states a *necessity*. The doctrine (ADR-0021 §3) only
    requires that promotion be a function of **coherence with the reviewed field**.
-   For the subclass of claims that are *deductively entailed* by already-`COHERENT`
-   facts, the entailment proof **is** that coherence judgment and can be
-   machine-certified by CORE's sound deductive engine — no human re-deciding
-   settled logic.
+   For the subclass of claims that are *deductively entailed* by an
+   already-`COHERENT` premise set, the entailment proof **is** that coherence
+   judgment and can be machine-certified by CORE's sound deductive engine — no
+   human re-deciding settled logic.
 2. **Code honesty (`teaching/epistemic.py`).** The enum docstring claimed
    transitions are *"computed from coherence with the existing reviewed field."*
    They are not computed; `review_correction` carries the status as a
@@ -106,7 +106,7 @@ reading**, and the design must put the whole defensive burden there:
    `COHERENT` in the vault — verified by status, not assumed. A premise that is
    `SPECULATIVE`/`CONTESTED` must make the proof inadmissible for promotion.
 3. **Premise selection / closure.** The premise set must be explicit and
-   bounded; the proof must not silently pull in unreviewed facts.
+   bounded; the proof must not silently pull in unreviewed claims.
 
 **Design consequence:** proof-carrying promotion does **not** remove human
 review — it *relocates* it. The human (or a separately-ratified structural

docs/position_paper.md

@@ -138,13 +138,13 @@ is retained for audit and revision, never as a promotion signal.
 
 That judgment is curator-mediated today, and for most corrections it must be. The
 fallible step is not the logic but the *reading* — translating a natural-language
-claim into the field's propositional form, and selecting which reviewed facts it
+claim into the field's propositional form, and selecting which reviewed claims it
 bears on. A sound inference over a misread premise is a sound proof of the wrong
 thing, so a human certifies the reading before a correction enters the reviewed
 structure.
 
 One subclass is different in principle. A claim that is *deductively entailed* by
-facts already marked `COHERENT` is not new information and is not the system's own
+claims already marked `COHERENT` is not new information and is not the system's own
 opinion — it makes explicit what the reviewed field already contains. For that
 subclass the entailment proof *is* the coherence judgment, and CORE's sound,
 independently-checked deductive engine (`deductive_logic_v1`, §4) can certify it