fix(hooks): remove interim truncation caps — reducer is the size gate

hooks/ways/check-bash-pre.sh

@@ -4,26 +4,16 @@
 # The ways binary handles: command pattern matching, semantic scoring,
 # check curve scoring, session state, and content output.
 #
-# The command is truncated to its semantic prefix before being passed to
-# `ways scan command`. Heredoc bodies (`gh pr create --body "$(cat <<EOF…)"`),
-# JSON payloads (`curl -d '{…}'`), and other large argument bodies carry
-# no signal for "what kind of command is this" — the program name and
-# first few args do. The MiniLM embedding models cap at ~128 tokens of
-# position embeddings; queries past that abort the embedder (ggml
-# get_rows out-of-range). 256 chars ≈ 60 tokens, safely under the limit
-# with headroom for the description that gets appended downstream.
-#
-# This truncation feeds both the embed query *and* the regex matcher in
-# the ways binary. Every existing `commands:` pattern under hooks/ways/
-# matches on the program name + first arg (≤106 chars), so cropping at
-# 256 changes no current behavior. Future patterns that need to look
-# past char 256 of a bash command would be misusing this trigger
-# anyway — that signal belongs in `pattern:` against the description.
+# Size bounding for the embed query is the ways binary's responsibility
+# (ADR-130 sentence-salience reducer in scan/reduce.rs). This script
+# passes the full command through so the reducer can score the prose
+# distribution itself — pre-truncating here would starve the reducer
+# of the back half of any long input. The regex `commands:` matcher
+# also gets the full command, which is what ways with patterns like
+# `^(npm|cargo|gh) ` expect.
 
 source "$(dirname "$0")/require-ways.sh"
 
-readonly CMD_QUERY_MAX=256
-
 INPUT=$(cat)
 CMD=$(echo "$INPUT" | jq -r '.tool_input.command // empty')
 DESC=$(echo "$INPUT" | jq -r '.tool_input.description // empty' | tr '[:upper:]' '[:lower:]')
@@ -32,11 +22,9 @@ AGENT_ID=$(echo "$INPUT" | jq -r '.agent_id // empty')
 [[ -n "$AGENT_ID" ]] && export CLAUDE_AGENT_ID="$AGENT_ID"
 PROJECT_DIR="${CLAUDE_PROJECT_DIR:-$(echo "$INPUT" | jq -r '.cwd // empty')}"
 
-CMD_QUERY="${CMD:0:$CMD_QUERY_MAX}"
-
 export CLAUDE_PROJECT_DIR="${PROJECT_DIR}"
 "${HOME}/.claude/bin/ways" scan command \
-  --command "$CMD_QUERY" \
+  --command "$CMD" \
   --description "$DESC" \
   --session "$SESSION_ID" \
   --project "$PROJECT_DIR"

hooks/ways/check-prompt.sh

@@ -9,18 +9,13 @@
 # harness also injects structured content here: <task-notification>
 # blobs from completed background agents, <persisted-output> pointers
 # for tool results that exceed inline budget, and other system-reminder
-# envelopes. Any of those can run multiple KB, and embedding that
-# overruns the MiniLM model's position-embedding table (SIGABRT in
-# ggml_compute_forward_get_rows). Cap the embed query at 1024 chars
-# (~240 tokens — generous because real user prompts can legitimately
-# be paragraphs of context, unlike bash commands). Anything past 1024
-# in a prompt is system-injected envelope content that carries no
-# additional signal for matching the user's *intent* against ways.
+# envelopes. Size bounding for the embed query is the ways binary's
+# responsibility (ADR-130 sentence-salience reducer in scan/reduce.rs).
+# This script passes the full combined prompt+topics through so the
+# reducer can score sentence salience across the whole input.
 
 source "$(dirname "$0")/require-ways.sh"
 
-readonly PROMPT_QUERY_MAX=1024
-
 INPUT=$(cat)
 PROMPT=$(echo "$INPUT" | jq -r '.prompt // empty' | tr '[:upper:]' '[:lower:]')
 SESSION_ID=$(echo "$INPUT" | jq -r '.session_id // empty')
@@ -37,11 +32,8 @@ if [[ -f "$RESPONSE_STATE" ]]; then
   RESPONSE_TOPICS=$(jq -r '.topics // empty' "$RESPONSE_STATE" 2>/dev/null)
 fi
 
-# Combined context: user prompt + Claude's recent topics, capped.
-# RESPONSE_TOPICS is bounded by check-response.sh's extraction (~50 chars
-# of keywords) so the cap is effectively a guard on PROMPT itself.
+# Combined context: user prompt + Claude's recent topics.
 COMBINED="${PROMPT} ${RESPONSE_TOPICS}"
-COMBINED="${COMBINED:0:$PROMPT_QUERY_MAX}"
 
 export CLAUDE_PROJECT_DIR="${PROJECT_DIR}"
 "${HOME}/.claude/bin/ways" scan prompt \