feat(skills): author the Continuous Research skills, and make its roles honest

The boot log has said `continuous_research — 4 role skills bound, 8 unresolved`
on every start. Those eight roles ran without the instructions their template
promised them, silently: `skills_loader` reports the miss and carries on.

Worse than the missing files was what the prompts described. v1 told the
harvester to sweep "RSS feeds, GitHub trending, HN front page, YouTube /
podcast RSS" — none of which exist. `harvest.rs` searches arXiv and nothing
else. A role prompt describing a machine that was never built is the failure
this codebase keeps paying for, so v2 describes the machine that exists.

Roles now match the pipeline: paper_reader (the harvest already ran; read the
manifest and the papers), signal_ranker (unchanged in spirit), script_writer
(the podcast half, which had no role at all).

Seven skills authored under skills/research/, kebab-case to match the loader —
team_template_loader.rs:177-181 documents the snake_case/kebab-case trap that
already unbinds skills elsewhere:

  arxiv-daily                  what the harvest guarantees, so an agent does
                               NOT re-search arXiv and corrupt the seen-set
  paper-to-project-relevance   name a file or roadmap item, or say "no bearing"
  duplicate-detection          the seen-set catches identity; this catches the
                               same work under a different id
  signal-to-noise-ranking      novelty/relevance/depth, and the two biases to
                               resist (recency up, inconvenience down)
  executive-summary-writing    what it is / why it matters / what to do — decide,
                               do not hedge
  obsidian-vault-conventions   the vault is a human's live workspace; never
                               main, never reorganise, hash the body not the file
  podcast-dialogue-writing     write for someone on a treadmill; the 10-70 char
                               highlight bound is the API's, not a style rule

`web_fetch` dropped from mcp_bundles: runtime_provision.rs binds every mission
claw to `["clawmates_door"]` and never reads that field, so declaring it
instructed roles to use a tool that never arrived. The prompts say `curl` via
Bash, which is what they actually have.

Boot now reports `continuous_research — 11 role skills bound`, with no
unresolved clause. 344 tests pass.

Co-Authored-By: Claude Opus 5 <[email protected]>

skills/research/paper-to-project-relevance.md

@@ -0,0 +1,65 @@
+---
+name: paper-to-project-relevance
+description: How to judge whether a paper actually bears on one of the operator's projects, and how to say so concretely instead of vaguely.
+when_to_use: You are reading harvested papers in a Continuous Research mission and must decide which ones matter and why.
+tags: [research, judgement]
+---
+
+# Does this paper bear on our work?
+
+The value of a research digest is not that it lists papers. It is that it tells
+you which of *your* problems just got easier. A summary that could have been
+written without knowing anything about the operator's projects has failed, no
+matter how accurate it is.
+
+## The test
+
+For each paper, answer in this order. Stop at the first "no".
+
+1. **Is the claim real?** What did they actually measure, on what, against what
+   baseline? A 40% speedup on a benchmark nobody runs is not a 40% speedup.
+2. **Does it touch a problem we have?** Not a problem in the same *field* — a
+   problem in the project brief. "Vector search" is a field. "Our HNSW insert
+   loop is serial and the ROADMAP flags it as needing a correctness pass" is a
+   problem.
+3. **Would we do anything differently?** If the honest answer is "no, but it's
+   interesting", say exactly that. It is a legitimate outcome and it costs the
+   listener five seconds instead of two minutes.
+
+## Say the specific thing
+
+Weak, and useless to act on:
+
+> This work on graph pruning could be relevant to vector search performance.
+
+Strong, because it names the target and the change:
+
+> They prune by neighbour-degree rather than distance, which is the same
+> decision `clawhdf5-ann/src/hnsw.rs` makes in `prune_connections`. Ours is
+> already rayon-parallel per node, so this is a swap of the scoring function,
+> not a restructure — maybe 30 lines. Worth trying because our ROADMAP lists
+> recall regression at high M as an open question, and that is exactly what
+> their Table 3 measures.
+
+Name a file, a function, a roadmap item, or an open question. If you cannot name
+one, you have not established relevance — you have established topic overlap.
+
+## Honest negatives are the point
+
+Most papers will not bear on the work. Recording that plainly is what makes the
+positives trustworthy. One line is enough:
+
+> No bearing — single-GPU training throughput, we do no training.
+
+A run where everything is relevant is a run that stopped judging.
+
+## What NOT to do
+
+- **Do not infer relevance from keyword overlap.** Sharing the word "agent"
+  with our codebase is not a connection.
+- **Do not soften a negative into a maybe.** "Could potentially inform future
+  work" is how a digest becomes unreadable.
+- **Do not review from the title or the abstract's claims.** The note carries
+  the abstract; the paper is one `curl` away. What a paper *says* it does and
+  what its evaluation *shows* are routinely different, and catching that gap is
+  most of the value you add.