fix(skills,install): alias quoting, binding cascade, silent CLEAN, two eval defects

[elhoim]

7 commits, one per fix, so each can be cherry-picked independently. Each was verified to apply cleanly onto 47df8ee by itself.

Commit	Fix
383d20a	quote the launcher paths written into the shell alias
59e6627	stop DeployCore running bun install across every user skill
2cacf0f	cascade binding removal to a fixpoint
d4cbd2f	never print CLEAN off a decode that failed
911c78e	survive a malformed user sources.json
c6d1c0b	diagnose a legacy v1 suite instead of throwing a TypeError
23f7df8	compute pass^k as all-k-passed, not as the mean

---

Seven defects across the skills payload and the installer.

• install.sh:314 (and its mirror at skills/LifeOS/install/install.sh) — the alias body interpolated $LAUNCHER/$SYS_PROMPT unquoted. A $HOME with a space breaks the launcher; one with a quote corrupts the rc file. By that point the awk pass has already commented out the working old alias, so the failure leaves the user with no alias. Adds shq() and quotes both parse levels. Tested with paths containing a space, a single quote, and a $(…) substitution: all four round-trip identically and nothing executes.
• skills/Tldraw/Tools/Tldr.ts:272 — a single forward pass over bindings leaves the start binding dangling when the target shape is removed. Reproduced with the tool's own validate: INVALID: binding:shape-a-start: dangling binding before, valid after (3 records removed vs 4). Iterates to a fixpoint.
• skills/AudioEditor/Tools/GateRepair.ts:185 — the verification decode's exit status was unchecked, so on decode failure it printed CLEAN and exited 0, certifying a file it never scanned.
• skills/LocalIntelligence/Tools/UserSources.ts:88 — unguarded JSON.parse; a trailing comma in the user's sources.json aborted the run after all eight fetchers had succeeded, so nothing was persisted.
• skills/Evals/Tools/EvalRunner.ts:133 — for (const c of suite.cases) throws a bare TypeError, because the only shipped suite (Suites/Regression/core-behaviors.yaml) is still the legacy v1 tasks: format. Callers that catch broadly turned that into a regression eval that silently never ran. I did not write a v1→v2 translation layer: UseCases/*.yaml define weighted graders and carry no prompt, so there is nothing for this single-shot runner to send, and guessing at the intended mapping seemed worse than a precise error. Before: TypeError: undefined is not an object (evaluating 'suite.cases'). After: a message naming the format, the 4 tasks, and the two ways forward.
• skills/Evals/Tools/EvalRunner.ts:149 — pass_to_k = passed/trials is a mean, not pass^k: 2-of-3 reported 67% where the true pass^k is 0, making a flaky case look mostly-passing.
• skills/LifeOS/Tools/DeployCore.ts:277 (and its mirror at LifeOS/Tools/DeployCore.ts) — walked all of configRoot/skills running bun install in every dir with a package.json, touching pre-existing user skills and double-installing the payload copy; failures land in blockers, so the documented offline install exits 1.

DeployCore.ts and install.sh are each shipped in two copies. Both copies were identical at 47df8ee and both are updated here — a payload file fixed in one copy and left broken in the other is its own defect.

---

Testing. Commands and output are in the per-file notes above. I did not do a fresh-system install verification (contributing step 3) — these are targeted fixes verified per-file, not an install run.