docs(claude): move architecture depth into skills

[prorochestvo]

Slims CLAUDE.md from 21133 to 11996 bytes (-43%), under the 12k budget its own routing block sets. No new skill was created — every block found a home in one of the five that already existed.

What moved where

Block	Destination
Source kind (BID/ASK/LAST), the options JSON column, cmd/doctor rulegen/audit, the Telegram proxy bypass and the dead HTTPS_PROXY/HTTP_PROXY/NO_PROXY variables	beacon-collection
Connection PRAGMAs, the BEGIN IMMEDIATE / deferred-read split and the SQLITE_BUSY story, the repository pattern, foreign-key cascade and RemoveRateSource, deploy-time schema reconciliation	beacon-storage
The initData HMAC scheme, --api-dsn, the PublicError contract and the three controller-error-test assertions	beacon-http-api
Pipeline procedure (plan / implement / gate / review / complete, plans/ layout, the four agents)	pipeline:working-agreement, via its delta form
Env var rows	internal/env.go, which already declares every name, format and consumer
Deployment mechanics	deploy/README.md

Each extended skill also got an extended description, so the new material has load triggers. A skill that gains content but not triggers loses it.

The beacon-http-api auth section previously read "The HMAC algorithm is in CLAUDE.md's Key Patterns" — that pointer would have dangled after this change, so the algorithm is now written out in the skill itself.

What stays in CLAUDE.md

The binary and layer map, the skills table, the routing block, ## Data & Privacy, ## Constraints, the gate, startup ordering, and a ## Tripwires section collecting every rule whose violation is silent — tiered reads, the weather_forecast_days 48 h sweep, the inner-mux auth mount, 404-not-403, r_* tags going straight to production.

Gate notes recorded

make lint and make lint-new each run two steps, golangci-lint run and scripts/lint-checks.sh — a green golangci-lint alone is not a green make lint. And on the pi5 make test dies compiling modernc.org/sqlite under -race; the rerun is go test -race -p 1.

Verification

All 227 backticked tokens were extracted from the old file. 225 appear verbatim in the new CLAUDE.md, one of the five beacon skills, pipeline:working-agreement, stack-go, or a repo file. The remaining two survive by meaning:

• /api/v1/me/... — kept as /api/v1/me/*, the same referent in the notation used everywhere else.
• /opt/beacon/artifacts/<VERSION_ID>/ — in deploy/README.md (layout tree plus the VERSION_ID definition) and literally in .github/workflows/release.yml.

Docs only; no code touched, so the gate is unaffected.

[prorochestvo]

Closing as salvaged rather than merged. 46b4698 (#142) carries the parts of this branch
that had not arrived by another route, taken verbatim rather than paraphrased.

First, the honest part: this branch merged clean against alpha as of 16ae37c. The two
conflicts it carries now — CLAUDE.md and beacon-storage/SKILL.md — are from merges I
landed on those same files afterwards (the long-range outlook, then #139, then #141). I did
not check for open work touching them before doing that.

What carried over

• beacon-collection — the ## Source rows: kind and options section; the
  ## Operator tooling: cmd/doctor section; and the reason the Telegram bot bypasses the
  proxy in code rather than by configuration — it is the channel that reports collection
  failures, so it must not share a failure mode with the thing it reports on. That reasoning
  existed nowhere else.
• beacon-http-api — the HMAC scheme with the key/message order called out as easy to
  invert; the --api-dsn rationale; and the whole ## Error rendering section carrying the
  PublicError contract and the three assertions a controller error test owes.
• beacon-storage — the repository-pattern paragraph, the source-deletion cascade
  subsection, and the deploy-time schema-reconciliation note.
• All four skill description blocks, which is the change with the longest reach here.
  The wider load triggers — proxyutil, domain.RateSourceOptions, BEACON_CHROMIUM_PATH,
  internal/tools/tgwebapp/initdata.go, internal.PublicError, RequireMigratedSchema,
  stubSQLiteDBThrough — decide whether the knowledge loads at all.

What did not, and why

The restructuring of CLAUDE.md into ## Architecture / ## Tripwires /
## Configuration. It is a bigger editorial call than a salvage should make on its own,
and the file is now 17823 — 2177 under its own 20k budget — so nothing forces it. Your
## Tripwires section was used as calibration while thinning: every fact it names is present
in the current file.

Also skipped: the passages this branch moves that had already been written by the outlook work
(#131) and by #141 — the long-range forecast section, the content-gated digest, the
weather_forecast_days-is-bounded subsection, the historical-migration-test note, and the
PRAGMA / write-lock depth.

The branch docs/slim-claude-md is left in place, so the restructure can be revisited as its
own change against current alpha rather than rebased through these conflicts.