docs: complete the open-source community standards

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -0,0 +1,104 @@
+name: Bug report
+description: Something in Traces behaves incorrectly.
+labels: [bug]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Two things make a Traces bug report actionable, and neither is obvious:
+
+        **Which WebMCP mode you were in.** The polyfill, the origin trial and the `webmcp-tools`
+        inspector are three different execution paths, and a bug in one often does not exist in the
+        others. The header pill in the app tells you: `live` means real WebMCP, `polyfill` means the
+        17 tools are not registered with a host.
+
+        **Which recording.** Every recording is synthetic and reproducible, so naming it makes the bug
+        reproducible too. If it only happens on a recording you made yourself, **describe it — do not
+        attach it.** A recording is a full reconstruction of a page and everything that was on it; see
+        [CODE_OF_CONDUCT.md](https://github.com/ribdsp/Traces/blob/main/CODE_OF_CONDUCT.md).
+
+        Found a security problem instead? Don't file it here — see
+        [SECURITY.md](https://github.com/ribdsp/Traces/blob/main/SECURITY.md).
+
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: What happened
+      description: What you saw, and what you expected instead.
+    validations:
+      required: true
+
+  - type: dropdown
+    id: surface
+    attributes:
+      label: Where
+      options:
+        - A WebMCP tool
+        - The replay player / timeline
+        - The findings panel (markers, hypotheses, tasks, report)
+        - The activity feed or undo
+        - bugbait (the demo checkout app)
+        - Build, install, or typecheck
+        - Somewhere else
+    validations:
+      required: true
+
+  - type: input
+    id: tool-name
+    attributes:
+      label: Tool name and arguments
+      description: If a tool was involved, name it and paste the arguments verbatim.
+      placeholder: 'bisect_state — { selector: "select[name=province]", predicate: { kind: "optionCount", equals: 0 } }'
+
+  - type: dropdown
+    id: webmcp-mode
+    attributes:
+      label: WebMCP mode
+      description: The header pill in the app says which one.
+      options:
+        - live — Chrome/Edge origin trial
+        - polyfill — the development fallback
+        - webmcp-tools browser extension
+        - Not applicable — the bug is in the UI
+    validations:
+      required: true
+
+  - type: input
+    id: agent-client
+    attributes:
+      label: Agent client
+      description: Which agent called the tool, if one did. "By hand from the inspector" is a valid answer.
+      placeholder: ChatGPT Desktop 1.2026.x / by hand from the inspector
+
+  - type: input
+    id: recording
+    attributes:
+      label: Recording
+      description: The slug from the picker, plus the timestamp in ms if the bug is at a particular moment.
+      placeholder: empty-province, around 28000ms
+
+  - type: textarea
+    id: steps
+    attributes:
+      label: Steps to reproduce
+      value: |
+        1.
+        2.
+        3.
+    validations:
+      required: true
+
+  - type: input
+    id: browser
+    attributes:
+      label: Browser and OS
+      placeholder: Chrome 149 / Windows 11
+    validations:
+      required: true
+
+  - type: textarea
+    id: console
+    attributes:
+      label: Console output
+      description: Errors from the devtools console, if there were any.
+      render: text

.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,17 @@
+blank_issues_enabled: false
+contact_links:
+  - name: Security vulnerability
+    url: https://github.com/ribdsp/Traces/security/advisories/new
+    about: Report privately. Please don't open a public issue for a security problem.
+  - name: The 17-tool contract
+    url: https://github.com/ribdsp/Traces/blob/main/docs/tools.md
+    about: Every tool's arguments, response shape, budget and failure modes. Read this before proposing an 18th.
+  - name: Architecture and threat model
+    url: https://github.com/ribdsp/Traces/blob/main/docs/architecture.md
+    about: How the pieces fit, and which boundaries are deliberate.
+  - name: The WebMCP specification
+    url: https://github.com/webmachinelearning/webmcp
+    about: Questions about WebMCP itself — including the missing "image" content type — belong upstream.
+  - name: The submission this was built for
+    url: https://devpost.com/software/traces-3snwtz
+    about: Traces was built for The WebMCP Challenge. Context, video, and the write-up live here.

.github/ISSUE_TEMPLATE/tool_proposal.yml

@@ -0,0 +1,85 @@
+name: Tool proposal
+description: Propose an 18th WebMCP tool, or a change to one of the 17.
+labels: [tool proposal]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        The tool surface is the interface a model reads, so it is designed rather than accumulated. The
+        questions below are the ones a proposal has to answer — they map onto
+        [CONTRIBUTING.md § Adding a tool](https://github.com/ribdsp/Traces/blob/main/CONTRIBUTING.md#adding-a-tool), and the second
+        one is the one that decides most proposals.
+
+  - type: textarea
+    id: what
+    attributes:
+      label: What the tool does
+      description: One paragraph. Include the name you'd give it and the shape of what it returns.
+    validations:
+      required: true
+
+  - type: textarea
+    id: why-the-page
+    attributes:
+      label: Why the page must compute this, rather than a server
+      description: >
+        The point of WebMCP here is that the answer lives in live browser state — a replayed DOM,
+        computed styles, layout boxes, a binary search that has to re-render the page 17 times. If the
+        same answer could come from an API call over the recording JSON, it does not need to be a tool.
+    validations:
+      required: true
+
+  - type: textarea
+    id: prompt
+    attributes:
+      label: The prompt an agent would be answering
+      description: >
+        Written the way a user would phrase it, not the way the tool would. A tool that only works when
+        the caller already knows the right arguments is a tool a model will not reach for.
+      placeholder: "Why is the Pay button still disabled at the end of the recording?"
+    validations:
+      required: true
+
+  - type: input
+    id: budget
+    attributes:
+      label: Proposed response budget
+      description: >
+        Every response has one — lines, characters, entries, or matches — plus a `truncated` flag and a
+        sentence telling the agent what to do about it. Say which.
+      placeholder: 20 entries, truncated → "narrow the time range"
+    validations:
+      required: true
+
+  - type: textarea
+    id: failure-modes
+    attributes:
+      label: How it fails readably
+      description: >
+        Every rejection path returns a sentence an agent can correct itself from, not a thrown
+        exception. List the rejections you can foresee and what each one would say.
+    validations:
+      required: true
+
+  - type: checkboxes
+    id: boundaries
+    attributes:
+      label: Boundaries
+      description: >
+        These come from the [threat model](https://github.com/ribdsp/Traces/blob/main/docs/threat-model.md). A proposal that needs one
+        of them relaxed is a discussion about the threat model first, and a tool second.
+      options:
+        - label: Nothing the model supplies would be executed — no `eval`, no `new Function`, no compiled regex, no interpolated expression.
+          required: true
+        - label: It reads recorded state only. It does not need a real user's session.
+          required: true
+        - label: It does not return input values or request/response bodies. The recorder never captures those.
+          required: true
+
+  - type: textarea
+    id: alternatives
+    attributes:
+      label: What you tried with the existing 17
+      description: >
+        Which of the current tools you composed, and where the composition fell short. See
+        [docs/tools.md](https://github.com/ribdsp/Traces/blob/main/docs/tools.md).

.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,71 @@
+<!--
+Thanks for this. The checkboxes below aren't ceremony: the three rules are the ones a reviewer
+cannot verify by reading a diff, and the verification line is the one that catches "it compiles on
+my machine". Delete any section that genuinely doesn't apply.
+-->
+
+## What this changes
+
+<!-- One paragraph. What was wrong or missing, and what it does now. -->
+
+Closes #
+
+## Area touched
+
+<!-- The codebase is split by folder, not by feature. Tick the ones this PR lands in. -->
+
+- [ ] `lib/replay` — recording load, checkpoints, event digest
+- [ ] `lib/dom` — DOM compression, diff, layout
+- [ ] `lib/bisect` — binary search and predicate evaluation
+- [ ] `lib/report` — report building and verification
+- [ ] `lib/webmcp` — tool registration, schemas, blocking gates
+- [ ] `components/` — UI
+- [ ] `bugbait/` — the demo checkout app
+- [ ] Docs only
+
+PRs touching `lib/dom`, `lib/bisect` or `lib/webmcp` need a second reader — those are where bugs hide
+without announcing themselves.
+
+## The three rules
+
+<!-- From docs/threat-model.md. A PR that breaks one is rejected even if everything else is good. -->
+
+- [ ] **Nothing from the model is executed.** No `eval`, no `new Function`, no dynamic `import()` of a
+      model-supplied string, no compiled regex from model text. New predicate capability means a new
+      variant plus its evaluator, never an escape hatch.
+- [ ] **Every new or changed tool response has a budget**, a `truncated` flag, and a sentence telling
+      the agent what to do about it rather than only that something was cut.
+- [ ] **No real user recording is added.** Every recording here is synthetic, generated by `bugbait/`.
+
+## Verification
+
+```bash
+cd traces && npm install && npx tsc --noEmit && npm test
+```
+
+- [ ] `npx tsc --noEmit` is clean
+- [ ] `npm test` passes
+- [ ] If this touches `lib/dom`, `lib/bisect` or `lib/report`: tests added
+
+Paste the test summary line:
+
+```
+```
+
+## If this touches a tool
+
+- [ ] Called from the `webmcp-tools` inspector with hand-typed arguments
+- [ ] Called from a real agent, with a task phrased the way a user would phrase it
+- [ ] Failures come back as readable tool errors, not thrown exceptions
+- [ ] [docs/tools.md](https://github.com/ribdsp/Traces/blob/main/docs/tools.md) updated — the contract is documentation, not just code
+
+## Not done
+
+<!--
+The `compress-dom`, `bisect` and `evaluatePredicate` suites were written first, as specifications.
+Never make one pass by weakening an assertion, adding `.skip`, or deleting a case — a red test there
+means the implementation is wrong. `no-eval.test.ts` is a security boundary, not a lint rule.
+
+If you found a problem outside your area, say so here instead of fixing it: a patch that helpfully
+repairs a neighbouring module hands its maintainer a conflict to resolve blind.
+-->