Skillscript

Authors' instinct is compute first, output last — every language they
know puts return/print at the end. Perry's qwen authorability test
caught the gap: when finally prompted to add an output template, the
model placed it at the bottom (after default:). Pre-v0.19.8 this
parsed as a malformed target with a misdirecting error:

# Skill: fetch-page
# Vars: URL=""

get:
    shell(command="curl -s ${URL}") -> FetchedPage
default: get

OUTPUTS: ${FetchedPage}      ← misparsed as `target OUTPUTS with deps`

→ Tier-1 [missing-dependency] Target 'OUTPUTS' depends on '${FetchedPage}'…

The fix — Pin 4 applied uniformly

Pin 4 was already the disambiguation rule for the top region. v0.19.8
extends it to all positions:

A column-0 <name>: line is a TARGET only if it is followed by an
indented op-block. Otherwise it's template text — regardless of
position (top, bottom, or both).

Subsumes the original top-only rule. Position is now free. Bottom-placed
templates parse cleanly: OUTPUTS: ${FetchedPage} (no op-block follows)
reads as template prose.

Multi-region templates — explicit error

If template content appears BOTH before any target AND after all
targets, raise a parse error per Perry's recommendation:

"Skill has body-text-as-output template content in two places
(before targets AND after targets). Pick one location: put the
template either above all targets OR below them, not both.
Concatenating across regions silently would hide the structural
ambiguity."

Loud over silent-concat. Forces the author to pick one location.

What changed

• src/parser.ts — removed the inTemplateRegion permanent-exit
  state. Template capture now runs on every line and applies Pin 4
  uniformly. Added recordTemplate helper + multi-region tracking.
  Blank lines inside target body (Bug 15 whitespace) NOT captured
  as template — only blanks outside any target body. The v0.19.7
  orphan-indented-op guard still fires for indented lines past any
  target (preserves the silent-drop fix).

• tests/v0.19.4-output-template.test.ts — 8 regression tests:
  bottom template parsing + compile + execute, Perry's qwen
  OUTPUTS: case rendering clean, multi-region parse error,
  top-only / bottom-only individually, Bug 15 target-body blanks
  not contaminating template, full bundled-corpus regression.

1883 total tests (+8 net new).

Migration

No migration required. Top templates continue to work unchanged.
Authors may now write the template at the bottom of a skill as
their language instincts suggest.

Note: bdf760f4 (tier-3 advisory) deferred

Perry's companion finding — a tier-3 "no explicit output" advisory
that fires when a skill binds vars but writes no template / emit —
was deferred per her own sequencing note: "the template-anywhere
change reduces this one's urgency — once authors can put the
template where instinct says (the bottom), the 'forgot it entirely'
rate should drop. So I'd sequence this after the position fix
and re-measure." Watching the next dogfood rung before deciding.