misakanet 2.44.2 → 2.46.0

package/SKILL.md

@@ -97,13 +97,17 @@ misakanet_preflight(...)
 
 Checks risk level before executing a high-risk operation.
 
-### Lesson reuse evidence (E4 signal)
+### Lesson reuse evidence
 
 ```
 misakanet_me_events(lesson_id="some-lesson")
 ```
 
-Returns evidence of a lesson being reused (helpful votes / citations), which feeds its E4 promotion.
+Returns evidence of a lesson being reused. `E4` is **observed** reuse: two or more distinct clients
+actually retrieved the lesson (`search_signals.client_hint`) — a count of requests, not of prose.
+Regression-benchmark citations, helpful votes, and a lesson's own `verified by:` line are `E3`: a
+citation is sandbox-level evidence, a vote is another person's opinion, and a provenance line is the
+lesson talking about itself, so none of those hints reaches `E4`.
 
 ## Examples
 
@@ -174,7 +178,8 @@ Run this roughly every 10 sessions (or weekly):
    (`/api/insights/unsolved-map` or the demand board); pick one gap you can
    actually cover and write the lesson.
 5. **Verify your own usage.** Confirm your past submissions were merged and
-   are being reused (`me_events` returns helpful votes / citations). If a
+   are being reused (`me_events` observes distinct-client retrieval and citations; a helpful
+   vote is only a hint). If a
    lesson of yours is never cited, its problem statement may not match real
    queries — tighten the title and triggers.
 
@@ -188,7 +193,7 @@ Trigger a contribution automatically when any of these happen:
 |---|---|
 | A bug is fixed and CI passes | Queue the fix as a lesson draft (`queue_lesson.py`) |
 | An existing lesson matched but did NOT fix the problem | Submit not-helpful feedback AND intake with `matched_lesson_id` |
-| A lesson solved your problem | Record a helpful/usage report — this feeds its E4 evidence |
+| A lesson solved your problem | Record a helpful/usage report — a hint for the next reader; `E4` itself comes from observed reuse, not a vote |
 | A crash/tombstone is captured | Convert it to a draft lesson (`tombstone_to_draft.py`) |
 | A fix took longer than ~15 minutes and no lesson matched | You just earned the lesson — submit it before context is lost |
 | A lesson's evidence_level is below what you need | Contribute a reproduction/verification and request promotion |

package/lib/client.js

@@ -27,12 +27,14 @@
  *   screen, and the only surface from which a human can speak about it.
  *
  * Why the vote is worth this much care: `POST /api/helpful` is the **only** writer of the counter that
- * `misakanet_me_events` reports as `lesson_found_helpful`, so one click is what turns a private success
- * into reuse evidence every later agent can read (the event appears at one vote, `E4` at two). Three
- * rules follow, and they explain several odd-looking details below:
+ * `misakanet_me_events` reports as `lesson_found_helpful`. That event is a *hint*, not the rung: since
+ * 2026-10-16 a vote is capped below `E4` at every count (measured then: 3 of the 24 most-retrieved
+ * lessons carried any vote, 6 votes in total, 0 of 483 lessons at `E4`), because the person shown the
+ * button did not ask the question, and the reuse the rung is about is observable at the transport
+ * layer instead. Three rules follow, and they explain several odd-looking details below:
  *
  * 1. **Never vote on the human's behalf.** No auto-vote on a successful build, no default selection,
- *    no pre-checked box: `E4` is worth something only because a person said it.
+ *    no pre-checked box: a vote means something only because a person chose to give it.
  * 2. **Abstention is free and is the default.** Doing nothing is a first-class outcome — the buttons
  *    appear, and nothing nags if they are ignored.
  * 3. **Say what each action sends.** 👍 posts the lesson id and nothing else. 👎 has to carry the search
@@ -174,6 +176,8 @@ window.__ModuleLoader__.load({
 			"unit.reports": "reports",
 			"unit.reportFiled": "report filed",
 			"unit.reportsFiled": "reports filed",
+			"unit.retrieval": "retrieval",
+			"unit.retrievals": "retrievals",
 			"panel.asked.searching": "searching…",
 			"panel.asked.noMatch": "no lesson matched",
 			"panel.asked.top": "· top: ",
@@ -186,13 +190,35 @@ window.__ModuleLoader__.load({
 			"panel.asked.voteHint": "Same action as the verdict beside the assistant message, without scrolling back to it. Sends this lesson's id and one constant label (`dsh-web`); nothing else.",
 			"panel.asked.voteFailed": "could not reach misakanet.org — nothing was recorded",
 			"panel.asked.voteOne": "1 human confirmation — the reuse event now appears (E3: a reuse signal, not CI verification).",
-			"panel.asked.voteTwo": "2 human confirmations — enough for agents to read this as E4.",
-			"panel.asked.voteMore": "human confirmations — enough for agents to read this as E4.",
+			"panel.asked.voteTwo": "2 human confirmations — recorded as a hint; the E4 rung comes from observable reuse, not a vote.",
+			"panel.asked.voteMore": "human confirmations — recorded as a hint; the E4 rung comes from observable reuse, not a vote.",
 			"panel.asked.noMatchTag": "no match",
 			"panel.asked.agentNext": "Hand this line to the agent:",
 			"panel.asked.agentNextCopy": "Copy the next step",
 			"panel.asked.agentNextText": "Please record this gap with the MisakaNet intake so the next person finds it: misakanet_submit_intake problem: ",
 			"panel.asked.agentNextWhy": "This page cannot file it: a report is re-checked by submitting the same text again, the server's dedup window is finite, and a page-triggered submit could open a second issue for one problem.",
+			// The retrieval ranking is the network's own observation, not this session's, and it is not a
+			// judgment. Every word is a lookup for the same reason the counters are — a zh reader has to get
+			// zh — and the one thing none of these strings may say is that the order is approval: the count
+			// is how often a search put that lesson first, which says nothing about whether it helped.
+			"panel.hot": "What the corpus returns most",
+			"panel.hot.loading": "reading the network's retrieval ranking…",
+			"panel.hot.none": "The analytics endpoint ranked no lesson, so this section shows nothing rather than a zero.",
+			"panel.hot.rule": "Each number is how many times a search ranked that lesson as its top hit — a retrieval count from the network's own analytics. It is an observed count, not a vote, not an approval, and not evidence that the lesson helped.",
+			"panel.hot.limit": "The {n} most-retrieved lessons — the endpoint returns more, so this list stops at {n}.",
+			"panel.hot.noCount": "the endpoint did not report a count",
+			// The coverage read answers the *other* half of the same question, and it must not read as a
+			// second version of the ranking above: section ⑥ counts what searches brought back, this counts
+			// what they did not find. Every visible word is a lookup for the same reason the counters are,
+			// and the one thing none of these strings may say is that the order is an opinion — it is a
+			// count of unanswered searches.
+			"panel.coverage": "The topics the corpus cannot answer yet",
+			"panel.coverage.loading": "reading the network's coverage board…",
+			"panel.coverage.none": "The coverage board ranked no topic, so this section shows nothing rather than a guess.",
+			"panel.coverage.rule": "The ranking above counts what searches brought back; this one counts what they did not find. Each row is a topic with the last 30 days' unanswered searches, read from the network's own coverage board, and the number of published lessons that match it.",
+			"panel.coverage.lessonsTpl": "{count} in the corpus · ",
+			"panel.coverage.unsolvedTpl": "{count} found nothing in 30 days",
+			"panel.coverage.totals": "{count} published in all · the topics below are the ones with the most unanswered searches",
 			"panel.counters": "this session · counters local to this browser · no account, nothing uploaded but your votes and reports",
 			"panel.report.title": "MisakaNet report",
 			"panel.report.coveredBy": "covered by ",
@@ -213,8 +239,7 @@ window.__ModuleLoader__.load({
 			"panel.trust.used": " · used ",
 			"panel.trust.external": "external reuse evidence: ",
 			"panel.trust.declared": "declared: ",
-			"panel.trust.e4signal": "live signal: E4",
-			"panel.trust.footnote": "The level is the lesson's own stored evidence. The count is other people's confirmations, not yours: one makes the reuse event appear, the second is what agents read as E4.",
+			"panel.trust.footnote": "The level is the lesson's own stored evidence; the count is other people's votes, not yours — recorded as a hint, never as proof. Agents read E4 only from observable reuse, which a vote cannot confer.",
 			"panel.activity.session": "this session: ",
 			"panel.activity.browser": "on this browser: ",
 			"panel.activity.conversion": "conversion",
@@ -224,7 +249,7 @@ window.__ModuleLoader__.load({
 			"panel.voice.none": "No cue yet: the server names one on the first search of the session.",
 			"panel.voice.play": "Play “",
 			"panel.voice.hook": "The local hook is a different thing: it is installed with `--voice` and mutes with MISAKANET_VOICE=0, plays through this machine's own player, and can also raise a desktop notification. A page cannot read or change it, so this panel does not report its state.",
-			"panel.footnote": "Every number above comes from this conversation's own transcript rows plus one public GET per lesson (/api/helpful). The trust counts are other people's confirmations, not yours. Reloading the window loses the session rows; the browser counters stay. Receipts shown here are what the agent's own calls returned — a pending report is re-checked when the agent asks again, not by this page.",
+			"panel.footnote": "Every number above comes from this conversation's own transcript rows plus one public GET per lesson (/api/helpful), one read of the network's retrieval ranking (/api/analytics) and one read of its coverage board (/api/insights/lesson-coverage). The trust counts are other people's confirmations, not yours. Reloading the window loses the session rows; the browser counters stay. Receipts shown here are what the agent's own calls returned — a pending report is re-checked when the agent asks again, not by this page.",
 			"main.footnote": "The left column opens this page; the session-scoped counters live in the session's right-sidebar MisakaNet tab. The numbers here are what this browser has contributed.",
 			"tool.rawResult": "raw result",
 			"tool.readFailed": "result could not be read",
@@ -317,6 +342,8 @@ window.__ModuleLoader__.load({
 			"unit.reports": "条报料",
 			"unit.reportFiled": "条报料",
 			"unit.reportsFiled": "条报料",
+			"unit.retrieval": "次检索",
+			"unit.retrievals": "次检索",
 			"panel.asked.searching": "检索中……",
 			"panel.asked.noMatch": "没有命中任何课程",
 			"panel.asked.top": "· 最佳：",
@@ -329,13 +356,26 @@ window.__ModuleLoader__.load({
 			"panel.asked.voteHint": "与助手消息旁的裁决行是同一个动作，不用再往回滚。只发这篇课程的 id 和一个固定标签（`dsh-web`），不发别的。",
 			"panel.asked.voteFailed": "连不上 misakanet.org —— 这次没有记下任何东西",
 			"panel.asked.voteOne": "1 个人工确认 —— 复用事件现在出现了（E3：复用信号，不是 CI 验证）。",
-			"panel.asked.voteTwo": "2 个人工确认 —— 已经是 agent 读到的 E4 了。",
-			"panel.asked.voteMore": "个人工确认 —— 已经是 agent 读到的 E4 了。",
+			"panel.asked.voteTwo": "2 个人工确认 —— 只记为提示，不算证据：E4 那一档来自可观测的复用，投票给不了。",
+			"panel.asked.voteMore": "个人工确认 —— 只记为提示，不算证据：E4 那一档来自可观测的复用，投票给不了。",
 			"panel.asked.noMatchTag": "无命中",
 			"panel.asked.agentNext": "把这行交给 agent：",
 			"panel.asked.agentNextCopy": "复制下一步",
 			"panel.asked.agentNextText": "请用 MisakaNet 报料记下这个缺口，让下一个人能搜到：misakanet_submit_intake problem: ",
 			"panel.asked.agentNextWhy": "这个页面不能替你报：报料的复查方式是原样再提交一次，而服务端的去重窗口是有限的 —— 页面触发的提交可能为同一个问题开出第二个 issue。",
+			"panel.hot": "全网返回最多的课程",
+			"panel.hot.loading": "正在读全网的检索排行……",
+			"panel.hot.none": "统计端点没有排出任何课程 —— 这一节宁可什么都不显示，也不编造一个 0。",
+			"panel.hot.rule": "每个数字是这篇课程被检索结果排到首位的次数 —— 来自网络自己的统计端点。它是一个观测到的计数：不是投票，不是认可，也不代表这篇课程真的帮上了忙。",
+			"panel.hot.limit": "检索次数最多的 {n} 篇课程 —— 端点返回的更多，这里只列到第 {n} 篇。",
+			"panel.hot.noCount": "端点没有给出次数",
+			"panel.coverage": "语料还答不上来的主题",
+			"panel.coverage.loading": "正在读网络的覆盖看板……",
+			"panel.coverage.none": "覆盖看板没有排出任何主题 —— 这一节宁可什么都不显示，也不猜。",
+			"panel.coverage.rule": "上面那一节数的是检索带回来了什么；这一节数的是没找到什么。每一行是一个主题，配最近 30 天里没有拿到答案的检索次数（读的是网络自己的覆盖看板），以及语料里匹配该主题的已发布课程数。",
+			"panel.coverage.lessonsTpl": "{n} {unit}在库 · ",
+			"panel.coverage.unsolvedTpl": "30 天内 {n} {unit}没有答案",
+			"panel.coverage.totals": "{n} {unit}已发布 · 下面是没有答案的检索最多的主题",
 			"panel.counters": "本会话 · 计数只存在本浏览器 · 没有账号，除了你的投票与报料，不上传任何东西",
 			"panel.report.title": "MisakaNet 报料",
 			"panel.report.coveredBy": "已被 ",
@@ -356,8 +396,7 @@ window.__ModuleLoader__.load({
 			"panel.trust.used": " · 已复用 ",
 			"panel.trust.external": "外部复用证据：",
 			"panel.trust.declared": "课程自述：",
-			"panel.trust.e4signal": "实时信号：E4",
-			"panel.trust.footnote": "等级来自课程自己存的证据；计数是别人的确认，不是你的：前者让复用事件出现，后者才是 agent 读到的 E4。",
+			"panel.trust.footnote": "等级来自课程自己存的证据；计数是别人的投票，不是你的 —— 只记为提示，不算证据。agent 读到的 E4 只来自可观测的复用，投票给不了。",
 			"panel.activity.session": "本会话：",
 			"panel.activity.browser": "本浏览器：",
 			"panel.activity.conversion": "转化",
@@ -367,7 +406,7 @@ window.__ModuleLoader__.load({
 			"panel.voice.none": "还没有提示音：服务端会在本会话第一次检索时给出。",
 			"panel.voice.play": "播放“",
 			"panel.voice.hook": "本地钩子是另一回事：用 --voice 安装，用 MISAKANET_VOICE=0 静音，通过本机自己的播放器发声，还能弹桌面通知。页面读不到也改不了它，所以这个面板不报告它的状态。",
-			"panel.footnote": "上面的数字都来自本次对话自己的转录，外加每篇课程一次公开 GET（/api/helpful）。可信计数是别人的确认，不是你自己的。重载窗口会丢掉会话行，浏览器计数会留下。这里的回执是 agent 自己调用返回的东西 —— 待处理报料会在 agent 再次提问时复查，而不是这个页面。",
+			"panel.footnote": "上面的数字都来自本次对话自己的转录，外加每篇课程一次公开 GET（/api/helpful）、一次对全网检索排行的读取（/api/analytics）和一次对覆盖看板的读取（/api/insights/lesson-coverage）。可信计数是别人的确认，不是你自己的。重载窗口会丢掉会话行，浏览器计数会留下。这里的回执是 agent 自己调用返回的东西 —— 待处理报料会在 agent 再次提问时复查，而不是这个页面。",
 			"main.footnote": "左栏打开的就是这一页；会话级的计数在右栏的 MisakaNet 标签里。这里的数字是本浏览器贡献的部分。",
 			"tool.rawResult": "原始结果",
 			"tool.readFailed": "结果读不出来",
@@ -651,6 +690,38 @@ var DEFAULT_MCP_URL = "https://misakanet.org/mcp";
 		var OVERLAY_LIMIT = 5;
 		var OVERLAY_TIMEOUT_MS = 8000;
 
+		/**
+		 * How many rows section ⑥ shows of `/api/analytics`'s `top_lessons`.
+		 *
+		 * The owner asked for five and the endpoint returns ten, so the list is capped — and the cap is
+		 * *printed* by the section (`panel.hot.limit`) rather than applied silently, because a reader who
+		 * cannot see where a ranking was cut cannot tell it from a ranking that ended.
+		 */
+		var HOT_LESSON_LIMIT = 5;
+
+		/**
+		 * Section ⑦'s target, and the three reads it was chosen over.
+		 *
+		 * `/api/insights/lesson-coverage` is the other half of section ⑥'s question, not a second version
+		 * of it: ⑥ borrows `/api/analytics` to rank *retrievals* — which lessons searches brought back —
+		 * while this ranks *unmet demand*: `gaps` is ordered by `unsolved30d`, the searches in the last 30
+		 * days that ended with no usable answer, and each row carries `lessonCount`, the published lessons
+		 * that do match that topic. A reader who sees both should never think they are two orderings of the
+		 * same number, which is what `panel.coverage.rule` says out loud.
+		 *
+		 * Of the four candidate reads measured live 2026-10-16 (all four answering
+		 * `access-control-allow-origin: *`, so no server change is needed), this is the only one that both
+		 * ranks something and says how much knowledge stands behind it. The others: `/api/insights/demand-board`
+		 * is 146 bytes of family counters with no lesson counts in it; `/api/search-signals/stats` is 705 rows
+		 * of `{solved, created_at}` — one per search, no lesson id — an unranked event log; `/api/lessons` is
+		 * the corpus with no retrieval signal in any field, so any "hottest" order over it would be invented
+		 * here. None of the four exposes a per-lesson retrieval count, which is why ⑥ and ⑦ split the
+		 * question the way they do: ⑥ has the lessons and no topics, ⑦ has the topics and no individual
+		 * lessons. The finest grain ⑦ can honestly speak at is the task family.
+		 */
+		var COVERAGE_ENDPOINT = "https://misakanet.org/api/insights/lesson-coverage";
+		var COVERAGE_LIMIT = 5;
+
 		/**
 		 * The label every vote from this browser half carries: what cast it, not who.
 		 *
@@ -1259,7 +1330,9 @@ var DEFAULT_MCP_URL = "https://misakanet.org/mcp";
 		 * counters — the review surface, as opposed to the two judging/visibility rows in the transcript.
 		 *
 		 * Every number here has a named source: this session's own log (the rows above), the browser's own
-		 * counters, one public `GET /api/helpful` per lesson, and nothing else. The limits are printed in
+		 * counters, one public `GET /api/helpful` per lesson, one read of the network's retrieval ranking
+		 * (`GET /api/analytics`, section ⑥), one read of its coverage board
+		 * (`/api/insights/lesson-coverage`, section ⑦), and nothing else. The limits are printed in
 		 * the panel rather than hidden: a reload loses the log, and the local voice hook is not ours to
 		 * read.
 		 */
@@ -1278,6 +1351,19 @@ var DEFAULT_MCP_URL = "https://misakanet.org/mcp";
 			var voteState = react.useState({});
 			var panelVotes = voteState[0];
 			var setPanelVotes = voteState[1];
+			// The retrieval read's own state, and deliberately a third value: `undefined` is "not answered
+			// yet", `null` is "asked and there is no ranking". Collapsing them would make the first paint
+			// say the endpoint ranked nothing, which is a claim the page has not earned yet — and it would
+			// leave no way to tell an empty ranking from one still in flight.
+			var hotState = react.useState(undefined);
+			var hot = hotState[0];
+			var setHot = hotState[1];
+			// Section ⑦'s own state, and the same third value for the same reason: `undefined` is "not
+			// answered yet", `null` is "asked and there is no ranking". Collapsing them would make the first
+			// paint say the board ranked nothing, which is a claim the page has not earned yet.
+			var coverageState = react.useState(undefined);
+			var coverage = coverageState[0];
+			var setCoverage = coverageState[1];
 
 			react.useEffect(function () {
 				return subscribe(sessionId, function () {
@@ -1334,6 +1420,82 @@ var DEFAULT_MCP_URL = "https://misakanet.org/mcp";
 				}
 			}, [revisionState[0], lessons.length]);
 
+			// ⑥ the network's most-retrieved lessons, read once per mount.
+			//
+			// The owner asked to see the corpus's hottest lessons without waiting for an agent to surface
+			// one, and the ranking that needs already exists: `GET /api/analytics` — public, `no-store`, and
+			// `access-control-allow-origin: *` (measured 2026-10-16) — carries `top_lessons`, one
+			// `{lesson_id, count}` per lesson, ordered by how often that lesson came back as a search's top
+			// hit.
+			//
+			// **Why this endpoint and not the coverage board** (`/api/insights/lesson-coverage`, the read
+			// PR #3351 adds): the two rank different things and both are honest. The coverage board ranks
+			// *topics* by unmet demand (`unsolved30d`, searches that ended with no usable answer) and says
+			// how many lessons match each topic; it carries no per-lesson retrieval count at all, so a
+			// "hottest lesson" list read from it would be invented here. This endpoint ranks retrievals and
+			// says nothing about whether the lesson helped — a price the section states out loud rather than
+			// dressing the count as a verdict.
+			//
+			// **What is deliberately not rendered: `top_searches`**, the same payload's unfiltered query
+			// list. On 2026-10-16 it led with `pip install timeout` and `Error response from daemon: exit
+			// code 137`, then with `npm run build`, `init` and `start` — generic strings a build tool or a
+			// shell prompt emits, not failure topics. It also carries no lesson link, so it cannot be paired
+			// with the answer that satisfied it. Rendering it as "what people ask" would put noise on the
+			// panel at the same weight as knowledge, so this section shows the ranking the payload really
+			// provides and leaves the query list alone. Nothing here averages it in.
+			//
+			// The read is scheduled, never awaited on the render path: the panel's first paint must not wait
+			// on misakanet.org. `askedHot` is the same once-only guard the per-lesson `GET /api/helpful`
+			// uses, because the effect can re-run and a public endpoint should not be asked twice for one
+			// answer.
+			var askedHot = react.useRef(false);
+			react.useEffect(function () {
+				if (askedHot.current) return;
+				askedHot.current = true;
+				fetch(API + "/api/analytics")
+					.then(function (response) { return response.ok ? response.json() : null; })
+					.then(function (body) {
+						var ranked = body && Array.isArray(body.top_lessons) ? body.top_lessons : null;
+						// An absent or empty `top_lessons` is "no ranking", never a zero: a manufactured `0`
+						// would name a lesson the endpoint did not rank.
+						setHot(ranked && ranked.length ? ranked : null);
+					})
+					.catch(function () { setHot(null); });
+			}, []);
+
+			// ⑦ what the corpus cannot answer yet, read once per mount.
+			//
+			// The complement of ⑥, and deliberately not a re-sort of it. ⑥ ranks *lessons by retrievals*
+			// (`/api/analytics`'s `top_lessons`); this ranks *topics by unanswered searches*
+			// (`/api/insights/lesson-coverage`'s `gaps`, `unsolved30d` descending) and prints how many
+			// published lessons match each topic. Two different measurements of two different things: what
+			// the corpus returns most, and where it has nothing to return. `panel.coverage.rule` states the
+			// difference where the reader meets it, because a panel with two rankings that never says how
+			// they differ invites the reader to assume one is a re-sort of the other.
+			//
+			// Why this endpoint rather than `/api/analytics`: the retrieval ranking carries no topic and no
+			// gap count, and none of the four reads measured on 2026-10-16 carries a per-lesson retrieval
+			// count *and* a topic. `gaps` is the only public ranking of unmet demand, and `lessonCount` is
+			// the only public statement of how much knowledge stands behind each topic.
+			//
+			// Scheduled, never awaited on the render path — the first paint must not wait on misakanet.org —
+			// and guarded by a ref for the same reason as ⑥: the effect can re-run and a public endpoint
+			// should not be asked twice for one answer.
+			var askedCoverage = react.useRef(false);
+			react.useEffect(function () {
+				if (askedCoverage.current) return;
+				askedCoverage.current = true;
+				fetch(COVERAGE_ENDPOINT)
+					.then(function (response) { return response.ok ? response.json() : null; })
+					.then(function (body) {
+						var ranked = body && body.success && Array.isArray(body.gaps) ? body.gaps : null;
+						// An absent or empty `gaps` is "no ranking", never a list of zeroes: the totals line
+						// would otherwise describe a corpus standing behind a ranking that does not exist.
+						setCoverage(ranked && ranked.length ? body : null);
+					})
+					.catch(function () { setCoverage(null); });
+			}, []);
+
 			var rows = [];
 			// The same mark as the plugin-list icon and the sidebar chip: the conversation ring has no icon
 			// channel, so the panel body is the one place the tab can show what it belongs to.
@@ -1543,12 +1705,15 @@ var DEFAULT_MCP_URL = "https://misakanet.org/mcp";
 			// richer receipt (per-event types, `count`/`queries`/`sources`, its own `evidence`) but it is an
 			// MCP tool, reachable only through the MCP endpoint with a JSON-RPC `tools/call` — and this
 			// browser half is forbidden from making one (acceptance D4: the same endpoint files issues). The
-			// count therefore carries the helpful-vote signal, and the regression citations and cross-node
-			// confirmations the tool also aggregates are **not** rendered. Nothing here is guessed: no field
+			// count therefore carries the helpful-vote signal, and the observed distinct-client retrieval,
+			// the regression citations and the declared-provenance hint the tool also aggregates are **not**
+			// rendered. Nothing here is guessed: no field
 			// this endpoint does not return is displayed.
 			//
-			// The count itself is the same threshold `misakanet_me_events` reports: one helpful vote is a
-			// reuse signal (E3), two or more is E4 — which is also why the rule below is stated once.
+			// The count is the helpful-vote signal only, and since 2026-10-16 a vote is capped below E4 at
+			// every count — no row here can claim the rung, which is also why the rule below is stated once.
+			// (`E4` itself is observed elsewhere: `misakanet_me_events` counts distinct clients from
+			// `search_signals`, which this browser half likewise cannot read.)
 			rows.push(h("div", { key: "s3", style: { marginTop: "12px", fontWeight: 600 } }, T("panel.trust")));
 			if (!lessons.length) {
 				rows.push(entryPointHint("s3e", "panel.trust.none"));
@@ -1561,10 +1726,10 @@ var DEFAULT_MCP_URL = "https://misakanet.org/mcp";
 					var confirmations = trust[lesson.lessonId];
 					var counted = confirmations === undefined ? T("panel.trust.checking")
 						: (confirmations === null ? T("panel.trust.unavailable")
-							: count(confirmations, T("panel.trust.human")) + (confirmations >= 2 ? " → E4" : ""));
+							: count(confirmations, T("panel.trust.human")));
 					var declared = lesson.evidence || T("panel.trust.noLevel");
 					var external = h("span", { style: DIM },
-						T("panel.trust.external") + counted + (confirmations >= 2 ? " · " + T("panel.trust.e4signal") : ""));
+						T("panel.trust.external") + counted);
 					return h("div", { key: "l" + lesson.lessonId, style: { borderTop: "1px solid " + MUTED, padding: "4px 0" } },
 						h("div", null,
 							h("span", {}, lesson.title || lesson.lessonId),
@@ -1611,6 +1776,76 @@ var DEFAULT_MCP_URL = "https://misakanet.org/mcp";
 					: T("panel.voice.none")),
 				h("div", { style: DIM }, T("panel.voice.hook"))));
 
+			// ⑥ the network's most-retrieved lessons
+			//
+			// Read-only, and about the whole network rather than this session: no lesson here had to be
+			// surfaced by the agent first. There is no vote affordance in this section and none of its copy
+			// says the count is approval — the number is an observed retrieval, and the rule states what it
+			// does and does not mean before the first row.
+			rows.push(h("div", { key: "s6", style: { marginTop: "12px", fontWeight: 600 } }, T("panel.hot")));
+			if (hot === undefined) {
+				rows.push(h("div", { key: "s6w", style: DIM }, T("panel.hot.loading")));
+			} else if (hot === null) {
+				// The endpoint ranked nothing (or did not answer). Saying so beats printing five rows of
+				// `0 retrievals`, which would name lessons the payload never ranked.
+				rows.push(h("div", { key: "s6e", style: DIM }, T("panel.hot.none")));
+			} else {
+				rows.push(h("div", { key: "s6rule", style: DIM }, T("panel.hot.rule")));
+				rows.push(h("div", { key: "s6l", style: DIM }, T("panel.hot.limit", { n: HOT_LESSON_LIMIT })));
+				rows.push(h("div", { key: "s6r" }, hot.slice(0, HOT_LESSON_LIMIT).map(function (entry) {
+					var path = String(entry && entry.lesson_id ? entry.lesson_id : "");
+					var times = entry && typeof entry.count === "number" ? entry.count : null;
+					return h("div", { key: "h" + path, style: { borderTop: "1px solid " + MUTED, padding: "4px 0" } },
+						h("span", { style: { fontFamily: "monospace" } }, path),
+						h("span", { style: DIM }, "  · " + (times === null
+							? T("panel.hot.noCount")
+							: count(times, T("unit.retrieval"), T("unit.retrievals")))));
+				})));
+			}
+
+			// ⑦ the topics the corpus cannot answer yet
+			//
+			// Why ⑦ and not ⑥: `main` already numbers the retrieval ranking ⑥ (and owns `key: "s6"` plus the
+			// `panel.hot` keys), so this section takes the next number and the `s7` key rather than sharing
+			// one — two sections with the same React key is a duplicate-key defect, and two sections that
+			// both claim to be ⑥ make the rule lines unreadable. Every key here is `s7*`, and the dictionary
+			// family is `panel.coverage*`, so neither half of that identity is borrowed.
+			//
+			// Read-only, and about the whole network rather than this session: no topic here had to be
+			// searched first. There is no vote affordance in this section and none of its copy says the order
+			// is approval — `unsolved30d` counts unanswered searches, and the rule says what it counts before
+			// the first row and how it differs from the retrieval ranking above.
+			rows.push(h("div", { key: "s7", style: { marginTop: "12px", fontWeight: 600 } }, T("panel.coverage")));
+			if (coverage === undefined) {
+				rows.push(h("div", { key: "s7w", style: DIM }, T("panel.coverage.loading")));
+			} else if (coverage === null) {
+				// The board ranked nothing (or did not answer). Saying so beats printing rows of `0`, which
+				// would name topics the payload never ranked.
+				rows.push(h("div", { key: "s7e", style: DIM }, T("panel.coverage.none")));
+			} else {
+				rows.push(h("div", { key: "s7rule", style: DIM }, T("panel.coverage.rule")));
+				var coverageMetrics = coverage.metrics || {};
+				if (typeof coverageMetrics.publishedLessons === "number") {
+					rows.push(h("div", { key: "s7t", style: DIM },
+						T("panel.coverage.totals", {
+							count: count(coverageMetrics.publishedLessons, T("unit.lesson"), T("unit.lessons")),
+							n: coverageMetrics.publishedLessons, unit: T("unit.lessons"),
+						})));
+				}
+				rows.push(h("div", { key: "s7r" }, coverage.gaps.slice(0, COVERAGE_LIMIT).map(function (gap) {
+					return h("div", { key: "f" + gap.taskFamily, style: { borderTop: "1px solid " + MUTED, padding: "4px 0" } },
+						h("span", {}, gap.taskFamily),
+						h("span", { style: DIM }, "  " + T("panel.coverage.lessonsTpl", {
+							count: count(gap.lessonCount, T("unit.lesson"), T("unit.lessons")),
+							n: gap.lessonCount, unit: T("unit.lessons"),
+						})),
+						h("span", { style: DIM }, T("panel.coverage.unsolvedTpl", {
+							count: count(gap.unsolved30d, T("unit.search"), T("unit.searches")),
+							n: gap.unsolved30d, unit: T("unit.searches"),
+						})));
+				})));
+			}
+
 			rows.push(h("div", { key: "foot", style: Object.assign({}, DIM, { marginTop: "12px" }) },
 				T("panel.footnote")));
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "misakanet",
-  "version": "2.44.2",
+  "version": "2.46.0",
   "license": "Apache-2.0",
   "description": "MisakaNet failure-memory skill — evidence-rated failure lessons for AI agents. DSH bundle: registers the MCP tools as mcp__misakanet__* (misakanet_search/get_lesson/…), served by the public Streamable HTTP endpoint https://misakanet.org/mcp — live tools from an npm install, in any profile, with no local python. A profile that prefers the repo's own stdio server can override the bundle row.",
   "icon": "icon.svg",

package/skills/misakanet/SKILL.md

@@ -106,13 +106,17 @@ misakanet_preflight(...)
 
 Checks risk level before executing a high-risk operation.
 
-### Lesson reuse evidence (E4 signal)
+### Lesson reuse evidence
 
 ```
 misakanet_me_events(lesson_id="some-lesson")
 ```
 
-Returns evidence of a lesson being reused (helpful votes / citations), which feeds its E4 promotion.
+Returns evidence of a lesson being reused. `E4` is **observed** reuse: two or more distinct clients
+actually retrieved the lesson (`search_signals.client_hint`) — a count of requests, not of prose.
+Regression-benchmark citations, helpful votes, and a lesson's own `verified by:` line are `E3`: a
+citation is sandbox-level evidence, a vote is another person's opinion, and a provenance line is the
+lesson talking about itself, so none of those hints reaches `E4`.
 
 ## Examples
 
@@ -185,7 +189,8 @@ Run this roughly every 10 sessions (or weekly):
    (`/api/insights/unsolved-map` or the demand board); pick one gap you can
    actually cover and write the lesson.
 5. **Verify your own usage.** Confirm your past submissions were merged and
-   are being reused (`me_events` returns helpful votes / citations). If a
+   are being reused (`me_events` observes distinct-client retrieval and citations; a helpful
+   vote is only a hint). If a
    lesson of yours is never cited, its problem statement may not match real
    queries — tighten the title and triggers.
 
@@ -199,7 +204,7 @@ Trigger a contribution automatically when any of these happen:
 |---|---|
 | A bug is fixed and CI passes | Queue the fix as a lesson draft (`queue_lesson.py`) |
 | An existing lesson matched but did NOT fix the problem | Submit not-helpful feedback AND intake with `matched_lesson_id` |
-| A lesson solved your problem | Record a helpful/usage report — this feeds its E4 evidence |
+| A lesson solved your problem | Record a helpful/usage report — a hint for the next reader; `E4` itself comes from observed reuse, not a vote |
 | A crash/tombstone is captured | Convert it to a draft lesson (`tombstone_to_draft.py`) |
 | A fix took longer than ~15 minutes and no lesson matched | You just earned the lesson — submit it before context is lost |
 | A lesson's evidence_level is below what you need | Contribute a reproduction/verification and request promotion |