fix(skill): carry pitch-vs-pulse rules into the comparison synthesis path

Design-vs-implementation review found the comparison path missing two
reinforcements the single-topic path has: the windowed-claims rule (no
trend verbs from one 30-day window) and an explicit artifact gate. It
also pointed the beat at a "narrative section" the comparison template
doesn't have. The per-entity template block now carries an explicit
optional pitch-vs-pulse slot (omit entirely when silent - no
placeholder), and the comparison instruction names that slot, the
windowed rule, and the no-fetch-no-pitch gate. Instructions now sit at
the emission point, per the file's own v3.0.6/v3.0.7 lesson about
distance between instruction and output.

skills/last30days/SKILL.md

@@ -641,7 +641,7 @@ Topic A (the main topic, first in the vs-string) uses outer `--x-handle`, `--x-r
 
 **Then do WebSearch supplements** for: `{TOPIC_A} vs {TOPIC_B} comparison {YEAR}` and `{TOPIC_A} vs {TOPIC_B} which is better` — these catch rivalry articles that per-entity passes might not surface.
 
-**Use `RESOLVED_POSITIONING` per entity (Step 0.55 item 6) in two ways.** First, ground each entity's `What it is` cell in its CURRENT fetched pitch - describe the entity as it pitches itself today, never from memory. Second, if an entity's month of evidence directly bears on its pitch - SUPPORTS a specific claim, CUTS AGAINST one, or the conversation is squarely ABOUT the pitched ground - say so in PROSE inside that entity's narrative section, anchored to the real item with its engagement. When the pulse is orthogonal to the pitch (on-entity but about something the pitch doesn't speak to), say NOTHING about the pitch: omission is the correct output, and a manufactured connection is worse than silence. Match altitude: test SPECIFIC claims ("zero-config", "fastest", an uptime number) against specific threads; never grade a broad tagline ("financial infrastructure") against an individual thread - it is too broad to hit or miss.
+**Use `RESOLVED_POSITIONING` per entity (Step 0.55 item 6) in two ways.** First, ground each entity's `What it is` cell in its CURRENT fetched pitch - describe the entity as it pitches itself today, never from memory. Second, if an entity's month of evidence directly bears on its pitch - SUPPORTS a specific claim, CUTS AGAINST one, or the conversation is squarely ABOUT the pitched ground - say so in ONE prose sentence inside that entity's section of the comparison synthesis (right after the Community Sentiment line - the template marks the slot), anchored to the real item with its engagement. When the pulse is orthogonal to the pitch (on-entity but about something the pitch doesn't speak to), say NOTHING about the pitch: omission is the correct output, and a manufactured connection is worse than silence. Match altitude: test SPECIFIC claims ("zero-config", "fastest", an uptime number) against specific threads; never grade a broad tagline ("financial infrastructure") against an individual thread - it is too broad to hit or miss. Keep claims windowed - "this month's conversation" - never trend verbs like "losing the narrative" that one 30-day window cannot support. If positioning was not actually fetched this run for an entity, skip both uses for that entity - never supply a pitch from memory.
 
 **Skip the normal Step 1 below** - go directly to the comparison synthesis format (see "If QUERY_TYPE = COMPARISON" in the synthesis section).
 
@@ -1273,6 +1273,8 @@ Voice contract LAWs 1, 3, 5 apply to comparisons unchanged (no `Sources:` block,
 
 **Community Sentiment:** [Positive / Mixed / Negative / Enthusiastic / Security-concerned / etc.] ({N}+ mentions across {source list})
 
+[Optional pitch-vs-pulse sentence - ONLY if `RESOLVED_POSITIONING` was captured for this entity AND the month's evidence directly supports a specific claim, cuts against one, or is squarely about the pitched ground: one windowed prose sentence anchored to a real item with engagement. Otherwise omit entirely - silence, not a placeholder.]
+
 **Strengths (what people love)**
 - [Specific strength with `per <source>` attribution]
 - [Specific strength with `per <source>` attribution]