fix(cli): read piped stdin to EOF instead of the first line (#166)

[penta2himajin]

Summary

cat file | qwisp chat and qwisp chat < file answered using only the first line of the input — no error, no warning, exit 0 — while both the help text and the usage line advertise piping as an input path. main.swift used readLine() (one line) for every non-argv prompt, with no TTY check, so the interactive case (where one line is what the user means) and the piped case shared the same call. Now stdin is read to EOF when it is not a TTY; readLine() stays for the interactive case, where reading to EOF would block until ^D.

Found while building #162's A/B harness: a 126KB (~35K-token) file was ingested as 52 tokens and reported prompt 52 tok (94 tok/s) · gen 32 tok (96.4 tok/s) — a completely healthy-looking run that measured nothing. Same silent-failure class as #151.

Related issue

Closes #166

Verification

Gates run on a tree containing only this change (the in-progress #162 knob was stashed and both schemes rebuilt first, so the green below is evidence for this commit and not a superset):

• scripts/test_raw.sh → RAWTESTS 98/98 PASS
• scripts/test_completion.sh → COMPTEST 97/97 PASS + [stdin] piped multi-line read to EOF: 1-line=11 tok < 2-line=23 tok ok
• scripts/test_bench_batch.sh → BENCHBATCHTEST PASS

The new regression check is GPU-free (QWISP_FAKE=1 uses FakeBackend: no model weights, only the tokenizer this gate already requires). It asserts a strict increase in prompt tokens between one-line and two-line stdin rather than an exact count, so it does not encode the tokenizer's segmentation.

Verified RED before the fix (both arms reported 11 tok on the pre-fix binary) and GREEN after (11 < 23).

• Tests pass
• Zero warnings
• Related docs updated (n/a — help text already documented the intended behaviour; the code now matches it)

Notes for reviewer

• Behaviour change is confined to the piped/redirected case. Interactive qwisp chat with no argument still takes one line, deliberately.
• The prompt is trimmed of surrounding whitespace/newlines so a trailing newline in a piped file does not change tokenization.