Restructure + compress instruction set for a 200K-context budget

[xerhab]

Summary

Started as a prose-trim pass on the auto-loaded instruction set (CLAUDE.md + .claude/rules/*.md)
for a 1M-context model; that only bought 0.4%. Mid-PR the target changed to a 200K-context model,
which needed real structural changes, not just tighter wording:

1. Narrowed over-broad paths: matches. agent.md and dsh.md both listed agent/**, so
   editing any agent-side file - even a pure-qwen or pure-native one - loaded content that had
   nothing to do with it. Narrowed both to the specific files they actually govern.
2. Relocated settled ADR narrative out of the auto-loaded set. dsh.md/qwen.md were
   architecture-decision records for now-fully-shipped work; their rationale, spike history and open
   questions moved to new docs/dsh-adr.md / docs/qwen-adr.md (read on demand, never auto-loaded),
   leaving each rules file with only the operative invariants + a pointer to the ADR for "why".
3. A deep compression pass across every file, explicitly overriding the repo's own "never delete
   rationale" convention (user-authorized given the budget constraint): historical bug narration,
   rejected-alternative prose and worked-example arithmetic collapsed to the rule + a short reason.
   Every never/must/always rule, identifier, XERK-nnn tag and Tests: pointer survives -
   verified by diffing every backtick identifier old vs. new and restoring the ~39 dropped as a
   side effect (env vars, constants, test names).

Total: 579,198 -> 377,743 chars (-35%), every file under the 40,000-char cap.

Worst-case co-load (files that auto-load when touching the given path):

Touch	Before	After
agent/hub-agent.py	336K chars / ~84K tok / 16 files	222K chars / ~55K tok / 16 files
turma/server.js	252K / ~63K tok / 9 files	161K / ~40K tok / 9 files
sessions.html	189K / ~47K tok / 5 files	124K / ~31K tok / 5 files
board.js	176K / ~44K tok / 6 files	122K / ~30K tok / 6 files

hub-agent.py stays the worst case because it's a genuine monolith every other file's invariants
legitimately depend on - closing that further means splitting the code file itself, out of scope.

Test plan

• wc -m on every file - none at/over the 40,000-char cap (repo's own CI gate re-checks this).

• Every cross-reference between rules files (and into docs/dsh-adr.md/docs/qwen-adr.md)
  resolves to a real file.

• Every rules file's YAML frontmatter (paths: block) parses and is well-formed.

• Diffed every backtick-wrapped identifier (env vars, constants, function names, test names)
  across the whole old vs. new ruleset; the ~39 that had been dropped as a side effect of
  compression were individually restored into the file where their fact now lives.

• Read every diff by hand across both commits.

• node --test turma/tests/*.test.js agent/tests/*.test.js .github/scripts/tests/*.test.js —
  1695/1695 pass.

• python3 -m unittest discover -s agent/tests — 2194 tests, 1 pre-existing failure unrelated
  to this change (below).

• git status confirms the change is markdown-only — no source file is touched.

Pre-existing test failure, reported not fixed

test_qwen_guard.TestQwenGuardShimEndToEnd.test_replace_in_aws_is_denied fails on this dev host
only, and fails identically on origin/main — this PR changes no Python.

Cause: the qwen guard shim matches its credential globs as an expanduser'd prefix against a
realpath'd target. On this machine ~/.aws is a symlink to /mnt/c/Users/mhabeeb/.aws (WSL), so
the realpathed target escapes the ~/.aws/** prefix and the write is not denied. The sibling
~/.ssh case passes because that directory is real.

This is exactly the residual gap .claude/rules/qwen.md already records as accepted ("a symlinked
HOME could dodge a credential glob") — it is just not hypothetical on a WSL host whose ~/.aws
points at the Windows filesystem. CI is green because GitHub runners have no such symlink. Flagged
rather than fixed: it is a real, if low-severity, guard gap and out of scope for a docs change.