═══════════════════ HOW THIS FILE IS USED ═══════════════════

This is the package's history accumulator, and it is the ONLY file permitted to contain history. Every
other surface here states what is true now: the board is current truth and deletes what is resolved, a
planning surface carries what remains, and a document that describes a former state is corrected rather
than annotated. The tense check enforces that everywhere except here.

**A removal writes here BEFORE it happens.** The board tool refuses to drain an absorbed item, and refuses
to remove a field, without a reference naming the heading its extraction landed under — because a
current-truth surface deleted without extraction destroys the only record that a finding existed. The check
decides that the reference RESOLVES and never that the extraction is faithful: extraction is a compression,
so a text comparison would fail every correct one and pass a verbatim paste. Fidelity stays with the writer.

**What belongs in an entry:** the defect, how it was found, and the MECHANISM it revealed — written as a
shape rather than as an episode, so a later reader can recognize the class in a surface that has nothing
else in common with this one. **What does not:** an item's addressing, its courtesies, its restatement of
what another party said, and anything a rule, a check or a planning row already carries. A finding absorbed
into a check is already durable, and extracting it again is a second copy that can disagree with the first.

**Where the host declares its own history file, that file is the destination and this one is unused** — one
history home per project, so the two never compete for the same fact.

**A heading is a kebab sentence naming the CLASS**, because a closure cites it by name and the citation must
resolve.

**A class entry — one whose heading is the kebab sentence above, rather than an extracted item heading —
states what it does NOT cover, under the lead `BOUNDARY:`**, naming each neighboring class
and the repair that neighbor takes. The lead is structural rather than a phrasing, so a reader arriving
under a citation finds the routing without reading the entry whole, and a check reads the same lead a writer
does rather than matching prose. **A class with no neighbor yet writes `BOUNDARY: —`**, which states that
the question was asked and answered, where an absent lead makes an oversight indistinguishable from a class
that genuinely stands alone.

**An entry that is INSTANCES rather than a class says so, and a lead binds it only where that lead's question
is well-formed for an entry with no class.** The binding is per LEAD rather than per entry, because the two
leads ask different things. `BOUNDARY:` asks what this heading does NOT cover — well-formed for instances,
which exist precisely to route a reader onward, and answered by naming the classes each instance belongs to.
`POPULATION:` asks what the members of this CLASS are — malformed for an instances entry, which has no class
for members to belong to, so the question has no answer rather than an unwritten one.

**A blanket *carries no lead* is wrong for the same reason**: an instances entry carries a BOUNDARY, which
routes a reader onward, and no POPULATION, because it has no class to state. So the header declares the binding per lead and
a walk deriving its leads from this header derives their bindings with them, rather than carrying a list of
exempt entries that would go stale the first time either set moved.

**The leads that presuppose a CLASS, and so do not bind an INSTANCES entry:** `POPULATION:`

**That line is the DECLARATION and the paragraphs above it are its reason**, which is the difference between
a binding a mechanism reads and one only a reader does. The walk collects the leads named on it in the same
backticked form it already collects every declared lead, so a lead added to this header carries its binding
in the same edit and the check learns neither name. A binding stated only in prose is a contract clause no
mechanism can reach — the shape this surface exists to record — so stating it twice, once for the reader and
once in the form the walk reads, is the collapse being unavailable here rather than a duplication: the
reason and the operand answer different questions and neither derives the other.

**A class entry states its members under the lead `POPULATION:`, and every member carries one of two words,
because a member something has already hit and a member the class merely predicts read identically once they
sit in one list.** `OBSERVED` names a member the tree has actually produced — a collision that happened, a
divergence that fired, a verdict that moved. `PREDICTED` names a member derived from the class against a
surface where nothing has gone wrong yet. **A class whose count came back empty writes `POPULATION: —`**,
which states that the enumeration ran and found nothing. **A class nobody has enumerated writes
`POPULATION: UNMEASURED`**, which states that the count has not been taken.

**Those three are the whole vocabulary and an ABSENT lead is none of them, which is why absence fails rather
than reading as unmeasured.** An entry saying nothing and an entry declaring that nobody has counted are
different claims: the first is an oversight indistinguishable from a decision, the second is a seat stating
the debt where a reader and a count can both see it. The roster draws the identical line one surface over —
its declared word for an unbuilt half is a value in the cell rather than an empty cell — and the reason is
the same: **a declared gap is countable and drainable, while a missing one is invisible to everything except
the reader who happens to notice.** So `UNMEASURED` is the honest state of most classes here and carries no
stigma; what it does not permit is silence.

**The lead is structural for the same reason the boundary lead is**, and deliberately in the same shape: the
mechanism that reads a lead reads whichever one this header DECLARES, so a second declared lead is an operand
the existing shape already accepts rather than a new parser. A population stated as prose beside the entry is
readable and unjoinable, which is the failure this file collects — so the form that carries it is the one a
lead-reader can consume.

**The two words are not a confidence tier and the distinction is not about how sure the author is.** They
name what the ENTRY IS EVIDENCE OF: an observed member is evidence the class exists, and a predicted member
is evidence only of the derivation that produced it. A population of one observed member and four predicted
ones is a class with one instance and a hypothesis, and a reader tallying five members learns the opposite —
which is the transcribed-count failure arriving inside a population.

**The obligation runs on the PREDICTED word only, and stating it that way is what keeps the cost honest.** An
observed member is what an author reaches for anyway; a predicted one is the claim that needs marking,
because it is the one that reads as measured while resting on reasoning. **A predicted member that something
later hits is rewritten to OBSERVED in place** — that is a current-truth correction rather than history, and
the entry never records that it changed.

═══════════════════ ENTRIES ═══════════════════