Read 0291's fourth reversal condition as a price paid rather than as a trigger

docs/decisions/0291-the-provider-the-socket-waits-on.md

@@ -4,6 +4,8 @@ Date: 2026-09-04
 
 Status: accepted. Supersedes nothing. Superseded by nothing.
 
+Narrowed-by: 0351, on the fourth reversal condition, which is read as a price this board has already paid rather than as a state it can reach
+
 Issue: #291
 
 ## The decision

docs/decisions/0351-the-condition-that-counts-this-board.md

@@ -0,0 +1,154 @@
+# 0351. The reversal condition 0291 counts this board with
+
+Date: 2026-09-18
+
+Status: accepted. Supersedes nothing. Superseded by nothing.
+
+Narrows: 0291, on the fourth reversal condition, which is read as a price this board has already paid rather than as a state it can reach
+
+Issue: #351
+
+## The decision
+
+0291's fourth reversal condition - the count of open issues on this board that
+can be worked without a request reaching zero - is read as a price already paid
+and not as a trigger that can fire: the count has a floor of one, the member
+holding that floor is #98, and neither of that member's two exits is open, so no
+state of this board takes the count to zero. What the condition names is
+therefore the throughput the wait has been costing since 0291 landed, which is a
+price the record already weighed, and the three conditions beside it stand as
+the ways the choice is retaken.
+
+## Why the count has a floor of one
+
+The command is 0291's own, and what it answers moves every hour:
+
+    gh issue list --repo Flowfin/core --state open --limit 500 --json number,labels \
+      --jq '[.[] | select([.labels[].name] | map(startswith("blocked-on-")) | any | not)] | length'
+
+The number is not what this rests on. What it rests on is that #98 is in the set
+it counts and cannot leave it, and that holds at any number.
+
+**Its first exit is closing, and closing is downstream of the wait.** #98's
+done-condition is the first release and an operator running the probe against
+their own server, and the chain from there arrives at the issue 0291 was written
+for:
+
+    for n in 92 39 27 291; do printf '%s\t' "$n"; \
+      gh api "repos/Flowfin/core/issues/$n/comments?per_page=100" --paginate \
+        --jq '[.[] | .body | split("\n")[0] | select(startswith("Waiting for:"))] | last'; done
+    92      Waiting for: Flowfin/core#39
+    39      Waiting for: Flowfin/core#27
+    27      Waiting for: Flowfin/core#291
+    291     Waiting for: act - a pure-Rust crypto provider for `rustls` reaching a release the target leg compiles on every triple in 0113's set (upstream).
+
+So that member drains after the wait this condition exists to retake, never
+before it. A condition whose firing is downstream of the thing it would end is
+not a way out of it.
+
+**Its second exit is a `blocked-on-` label, and that move is argued shut on #98
+itself.** The label subtracts an issue from the count without closing it, it has
+been applied and removed twice, and the removal of 2026-08-22 gives its reason
+in the issue's own words:
+
+> Nothing holds it. It is the index, and an index that waits on the things it
+> indexes cannot be read first, which is the one job it has.
+
+    gh api repos/Flowfin/core/issues/98/timeline --paginate \
+      --jq '.[] | select(.event=="unlabeled" or .event=="labeled") | "\(.created_at) \(.event) \(.label.name)"'
+
+That reason is about what the index is for rather than about the state of the
+board on any day, so it does not expire with a milestone.
+
+Both exits shut, the floor is one, and zero is a state no arrangement of this
+board's open issues produces.
+
+## What moved between the reading and this record
+
+#351 derived the count as one with #98 as its only member, and the decision of
+2026-09-18 restates that figure. Read on 2026-09-18 the same command answers
+eight, and #98 is one of the eight. The difference changes nothing this record
+decides, because the argument above needs the floor and not the number: the two
+exits are shut at any count, and a count of eight is no closer to zero than a
+count of one.
+
+It is written here rather than left out because the count is the one figure in
+the argument that moves, and a reader who runs the command and gets a number
+neither the issue nor the decision names should find that difference already
+accounted for instead of concluding that one of the three is wrong.
+
+## What this does not decide
+
+- **Whether waiting is still right.** The choice 0291 holds is untouched. This
+  reads one of its four reversal conditions and takes no position on the other
+  three or on the decision they would reverse.
+- **Whether the condition should be dropped or rewritten.** Rewording what would
+  reverse a landed record is not one of the edits 0001 allows, so the condition
+  stands in 0291 exactly as written and this record says how it is read.
+- **Whether #98 should ever be closed or labelled.** Both of its exits are read
+  here as they stand today. Neither is argued for or against.
+- **The other three conditions.** A provider reaching a release the target leg
+  compiles everywhere, a runner image measured with a working C cross-toolchain,
+  and a triple leaving 0113's set are each a state of the world outside this
+  tracker, and each one still fires.
+- **Anything about the count's definition.** What subtracts an issue is a
+  `blocked-on-` label and nothing else, which is the rule the command encodes,
+  and this record neither widens nor narrows it.
+
+## Why this is written down
+
+Without it the reading lives in a comment on #291 and in the body of #351, and
+the next reader of 0291 meets a reversal condition that reads as live. Two
+things follow from that, and the second is the expensive one. A reader who takes
+it for a trigger waits for a state this board does not produce, which is the
+failure the condition was written against, arriving through the condition itself.
+And a reader who notices that it cannot fire re-derives this argument from the
+tracker - the chain behind #98, the label history, the reason of 2026-08-22 -
+which is an hour of reading that ends where this record already is.
+
+A comment is not a home for it either way. #291's done-condition is about a
+socket, a record and a green target leg, so the question cannot close with it,
+and what lives only in a comment is re-derived by whoever reads the record next.
+
+## Alternatives, and what each cost
+
+**Leave the condition unread.** It is the state that produced this issue: the
+reading existed on #291, was correctly declined there as a rule question, and
+had no home. It costs nothing today and costs the whole argument again at every
+reader, and the failure it leaves in place is somebody waiting for zero.
+
+**Rewrite the condition inside 0291 so it names a reachable state.** The cheapest
+thing to type and the one 0001 refuses: rewording is not among the edits the
+register allows, and a rule about what a reversal condition may say would be
+settled by landing a diff rather than by deciding it. It would also lose the
+price the original sentence records, which is the part that turned out to be
+true.
+
+**Supersede 0291.** It would carry the condition forward corrected, and it
+discards the decision, the costs it weighed, and the three conditions that do
+fire, in order to repair the one that does not. That is the price 0269 and 0313
+were both written to avoid paying for a single line.
+
+**Subtract #98 by labelling it.** This makes the count reachable at once and
+makes it dishonest: the label means an issue waits on another open issue on this
+board, #98 waits on nothing, and the board's index would be hidden from the
+count that decides whether the board has work. The reason was written on #98 on
+2026-08-22 and is quoted above rather than re-argued here.
+
+## What would reverse this
+
+#98 gains an exit this board reaches. Either its done-condition stops being
+downstream of the wait, or the reason of 2026-08-22 is overturned and the index
+is subtracted from the count some other way. Either one puts a reachable zero
+back and makes the condition a trigger again, and this record is then wrong in
+the direction that is safe to discover.
+
+The population the count is taken over changes - a second issue takes up
+permanent residence in it for a reason of its own, or what subtracts an issue
+stops being the `blocked-on-` prefix. The floor then has to be re-derived, and
+the reading above is about #98 alone rather than about whatever else is standing
+there.
+
+One of 0291's other three conditions fires. The wait ends on its own terms, the
+record naming the provider supersedes 0291, and this narrowing goes with it
+rather than outliving the sentence it reads.

docs/decisions/README.md

@@ -73,3 +73,4 @@ their wording, is under `## Decisions` in [CONTRIBUTING.md](../../CONTRIBUTING.m
 - [0292. The means the core speaks HTTP with, and what it costs](0292-the-means-the-core-speaks-http-with.md)
 - [0303. The licence the core is offered under, and the dependency set that follows from it](0303-the-licence-the-core-is-offered-under.md)
 - [0313. A dated re-reading of a landed record](0313-a-dated-re-reading-of-a-landed-record.md)
+- [0351. The reversal condition 0291 counts this board with](0351-the-condition-that-counts-this-board.md)