One prerequisite emitter, not four doors into NSpecIns

Filed by the agent that added two of the four doors, on the day it added them. Not a criticism of the increments — each was the right local change and each landed with its regression set green. The point is that the shape is now visible and will not be visible later.

The population, named rather than gestured at

compiler/pasparser_generic.inc, HEAD of 2026-08-30:

Three emitters feed them — EmitSpecDecl, EmitQualAliasDecl, EmitHoistedDecls — plus QualPush, which two of the three use for individual tokens and the third does not.

Why this is a design flaw and not tidying

Everything below is a rule that exists in one door and not the others, which is the definition of the second path that stays broken:

• The leading type keyword. Exactly one site emits it, with a correct and well-documented reason (the section loop has already exited, so a bare X = specialize Y<Z>; is not a declaration there). Nothing states the rule where the other three could see it, so the fifth door will get it wrong in whichever direction its author happens to test.
• The capacity check. EmitSpecDecl guards 8 + 2*nArgs; QualPush guards one token at a time; EmitHoistedDecls guards n + 4 before a bulk copy. Three encodings of one buffer's limit, and the buffer is sized MAX_NESTED_SPECS * (8 + 2*MAX_TEMPLATE_PARAMS) — a formula that mentions nested specs and now carries hoisted record bodies, which have nothing to do with MAX_TEMPLATE_PARAMS.
• The ordering rule. "Hoists, then qualified aliases, then nested specs, then self" is real and load-bearing — each names the one before it — and it exists only as the statement order inside one begin/end. It is not checked, not named, and not visible from the other three doors.
• The deferral trigger. (NSpecCount > 0) or (QNeedCount > 0) or hoistPending — three counters that mean the same thing ("something must be declared ahead of this"), each added by a different change, each with its own reset points. A fourth kind of prerequisite means a fourth counter and a fourth clause, and forgetting one is silent.

The reduction argument, which is what makes it worth doing

The proposal is a single QueuePrereq(kind, ...) / FlushPrereqs(at, needsTypeKeyword) pair: one buffer, one capacity rule, one ordering (queue order), one "is anything pending" predicate. Then:

• the leading-type decision becomes an argument at one call site instead of a fact three sites do not know;
• hoistPending / QNeedCount / NSpecCount collapse into PrereqCount, and the deferral condition stops growing a clause per feature;
• the next prerequisite kind is a new kind value rather than a fourth door;
• bug-p-a-delphi-mode-generic-argument-must-be-declared-before-the-template gets materially easier, because its fix is per-use anchoring — several splices at several points instead of one ordered run — and that is a change to the flush, not to four sites.

That last point is the honest reason this is ranked at all: it is a refactor with a named ticket it unblocks, not a cleanup for its own sake.

Timing — deliberately NOT taken when filed

Filed during a fleet-wide pause before a re-pin. A broad rewrite of a shared frontend file in the hour before a binary is blessed is the one window where this shape of change is wrong, and a partial normalisation is worse than either end state. Same call frankC made the same day on an until (depth = 0) or (CurTok.Kind = tkEOF) repeated at four sites: not landing the one-liner, because fixing one arm is how you get the arm that stays broken.

Gate

make compiler/pascal26 (the byte-identical self-host fixedpoint) + python3 tools/forwardlint.py (read the output) + the full named generic set: the 25 in test-core plus test_generic_qualified_arg{,_delphi} and test_generic_nested_type_as_argument, each diffed against its .expected or against pinned. test_generic_cycle_fail must still fail with its cycle diagnostic — it is the control for every change to the deferral path and the one this refactor is most likely to break. Track T sweeps the matrix.

2026-09-05 (frankA) — count RE-DERIVED and it holds at four; the line numbers do not

The body's table is dated 2026-08-30 and its four line numbers have all drifted. Re-derived at 3e25c7ae5 from the door predicate rather than the names:

grep -n 'NSpecInsCnt *:= *0'          compiler/*.inc
grep -n 'InsertTokens(.*NSpecIns'     compiler/*.inc

Four resets, four splices, one file, pairing exactly:

FOUR is correct — this ticket is not overselling itself. Worth stating explicitly because the sibling refactor [[refactor-p-three-hand-rolled-postfix-loops]] WAS underselling itself by 40% when I recounted it, and tonight frankB found a "twice" that was three times with the uncounted copy carrying two silent bugs. The instrument has now failed in both directions in one week, so a holding count is a result, not a formality. The greps above are the durable half; the line numbers will drift again.

2026-09-05 (frankA) — DONE, in two gated halves (101bf561f, f905ff471)

Four doors are one. NSpecInsCnt is touched only by BeginPrereqs and FlushPrereqs now; every filler appends through QualPush / PrereqPushRaw under one bound, PrereqReserve, against one named capacity, PREREQ_CAP — which the buffer's own declaration derives from, so the array and its guard cannot disagree.

Each of the four private facts the body listed, and where it went:

EmitSpecDecl went 42 lines to 21: it open-coded the same six-line slot fill ten times and reserved after the first, so a buffer filling mid-declaration could splice half an alias = specialize Name<args>;. It reserves the whole declaration up front now.

The gate, and why the fixedpoint is not it

This adds procedures, which changes declaration order, which is exactly what [[refactor-p-carve-out-paslexer-so-p-owns-its-lexer-too]] measured as the thing that moves the compiler's own bytes. So the sha is expected to move and says nothing. What must not move is the code it EMITS.

All 47 test_generic_* sources, compiled by the compiler from before step 1 and by the final one: 32 byte-identical binaries, 15 refusing with byte-identical diagnostics, 0 differing, 0 exit-code differences. Comparing diagnostics rather than just exit status is what makes the 15 negative tests carry signal instead of passing by failing.

Positive control on the same harness against the pinned compiler: 0 identical, 32 differing. test_generic_cycle_fail still fails with its cycle diagnostic naming both sides. forwardlint ok. gate.sh quick GREEN with the FPC seed canary concurrent — the canary matters here because two routines were forward-declared across their first call.

What it unblocks, which was the ranking argument

[[bug-p-a-delphi-mode-generic-argument-must-be-declared-before-the-template]] wants per-use anchoring — several splices at several points instead of one ordered run. That is now a change to FlushPrereqs and its call sites rather than to four independent doors. Not attempted here.

Log

• 2026-09-05 — resolved; this names the commit that carried the resolve, which is not always the one that carried the change — commit edf8afadf.