fix: TOML-escape iris token and base URL in docker entrypoint (COD-470)

[shivros]

Summary

Fixes a TOML escaping defect in docker-entrypoint.sh discovered while executing the deferred COD-467 Docker materialization acceptance check against merged main (a868849).

The defect: the entrypoint escaped RITE_IRIS_API_TOKEN SQL-style (' → '') inside a TOML literal string. TOML literal strings cannot contain single quotes and have no escape mechanism — a token like o'brien generated an unparseable rite.toml and a crashing container. base_url was also interpolated unescaped into a double-quoted basic string, breaking on " or \, and raw control characters are invalid in basic strings.

Reproduction on merged main:

RITE_IRIS_API_TOKEN=o'brien ./docker-entrypoint.sh true
# → api_token = 'o''brien'   ← invalid TOML
# tomllib.TOMLDecodeError: Expected newline or end of document (line 6)

Fix

• toml_escape() emits TOML basic strings: \\ and \" escaped, newline → \n, tab left raw (legal), every other control byte (0x01–0x1F) plus DEL → \uXXXX — applied to both api_token and base_url.
• GNU sed -z (NUL-delimited records; env values cannot contain NUL) so embedded newlines are matchable without a sentinel byte.
• Runtime image is Debian bookworm-slim (GNU sed), script stays POSIX-sh.

Test coverage (new: crates/rite-server/tests/docker_entrypoint.rs)

Runs the real entrypoint script and parses its output through rite's production parser (rite_server::load_config), asserting exact round-trip:

• plain / token-unset / blank base_url (no [sources.iris] emitted)
• o'brien, to"ken, back\slash, mixed delimiters
• embedded newline, CRLF, tab
• full sweep of every control byte 0x01–0x1F except tab/newline, plus DEL, in the token
• control bytes + quotes + backslash in base_url
• token value never echoed to stdout/stderr

Review panel

Both seats BLOCKed the first version (correctly — findings below were fixed in this head):

• STX (0x02) was used as a newline sentinel → collided with real STX bytes (both reviewers, MAJOR) → fixed: sed -z removes the sentinel entirely; STX regression test added.
• Only 0x01 of the control range was escaped; 0x03–0x1F and DEL leaked raw → invalid TOML (both reviewers, MAJOR) → fixed: complete \uXXXX table; full-sweep test added.
• Tests missed control/newline coverage in base_url (Reviewer A MINOR) → fixed: dedicated test.
• generated_token_is_never_logged didn't assert exit success (Reviewer A MINOR) → fixed.
• Round-trip failure messages dumped the generated TOML — potential token leak into test output (Reviewer A MINOR) → fixed: length-only messages.
• Invalid-UTF-8 env values (Reviewer A MAJOR): not fixable at this boundary — Rust's own env access is UTF-8-only (std::env::var fails on non-UTF-8), and rite's read_to_string startup path already requires UTF-8 configs, so non-UTF-8 values fail loudly at the same place they would have before. Documented in the script comment instead.

Verification

• cargo build --all-targets ✅
• cargo test --all-targets ✅ (36 tests / 6 suites, incl. 14 new)
• cargo clippy --all-targets -- -D warnings ✅
• cargo fmt --all -- --check ✅
• cargo run -p rite-codegen -- check ✅
• sh -n docker-entrypoint.sh ✅
• Direct materialization checks against merged main (defect) and this branch (fix) with tomllib + od byte-level inspection: raw SOH/STX/newline never appear on disk, only \uXXXX/\n escapes.

Closes COD-470 (discovered during the COD-467 acceptance check).

[shivros]

🤖 Automated Review Panel

Dual review of e7d7c22 vs main. Both seats BLOCKed the first head (7a79ffe) — findings applied and re-gated; this head is the corrected version.

Reviewer A — BLOCK → resolved

#	Sev	Finding	Resolution
1	MAJOR	tr '\n' '\002' sentinel collides with real STX bytes → emitted as \n, corrupts round-trip	Fixed: GNU sed -z (NUL-delimited records; env values can't contain NUL) — no sentinel byte. STX regression test added
2	MAJOR	Invalid-UTF-8 env bytes copied unchanged; read_to_string startup path can't load them	Documented, not fixable here: Rust's std::env::var is UTF-8-only and rite's startup already requires UTF-8 configs — non-UTF-8 values fail loudly at the same boundary as before this change. No behavior regression
3	MINOR	Coverage missed STX collision, remaining controls/DEL, invalid UTF-8, control/newline in base_url	Fixed: full 0x01–0x1F (+DEL) sweep test, STX test, control_chars_in_base_url_round_trip test
4	MINOR	generated_token_is_never_logged didn't assert exit success	Fixed: asserts status.success() first
5	MINOR	Round-trip failure messages interpolated the generated TOML (token could leak to test output)	Fixed: length-only failure messages
—	—	scratch_dir PID + atomic counter: race-free	No action needed

Reviewer B — BLOCK → resolved

#	Sev	Finding	Resolution
1	MAJOR	STX sentinel collision (converges with A#1)	Fixed as above
2	MAJOR	Only 0x01 escaped; raw 0x03–0x1F + DEL survive into basic strings → invalid TOML (verified 0x03, tab, 0x1f, 0x7f)	Fixed: complete \uXXXX escape table (tab left raw — legal in basic strings); sweep test proves all of them
3	MINOR	"BOTH fields" requirement not regression-protected (only token had control/newline coverage)	Fixed: dedicated base_url test with newline, quote, backslash, SOH
—	—	sed \x01 pattern behavior identical under LC_ALL=C and UTF-8 locales	No action needed

Panel outcome: all five blocking/major findings resolved in e7d7c22 (the UTF-8 item is a documented boundary, not a regression). Gates after fixes: build ✅, cargo test --all-targets 36/36 ✅ (14 new), clippy -D warnings ✅, fmt ✅, codegen check ✅, sh -n ✅.

Note: this panel ran on glm-5.1 children (cron pin differed from the usual gpt-5.6-terra runner seat). Both seats produced reproducible, verified findings; Reviewer A re-ran the failing materializations manually before blocking.

[shivros]

🤖 Review Panel Addendum — CI-caught defect fixed in 2bb0f57

CI on e7d7c22 failed the new entrypoint suite with a pattern invisible on the authoring host: backslash/newline/CRLF/control-sweep tests failed while quote/STX/plain tests passed.

Root cause: /bin/sh on GitHub's runner — and, more importantly, on the production runtime image (Debian bookworm) — is dash, whose builtin echo interprets \n, \\, \b, \f, \r in its arguments. The escaper produced correct TOML escape sequences; echo then mangled them (\\n in the token value became a literal newline on disk). The authoring host's /bin/sh is bash-as-sh, which doesn't interpret escapes — local gates were green while production would have been broken. The original entrypoint (echo with simple values) never hit this because unescaped values contain no backslash sequences.

Fix: emit all generated lines with printf (only the format string is escape-processed). 8 lines changed, no behavior change otherwise.

Verification:

• The fixed script was executed inside debian:bookworm-slim itself (the production base image): /bin/sh → dash, GNU sed 4.9. With RITE_IRIS_BASE_URL='https://x"a\b' and a multi-line quote/backslash token, the generated config is exactly base_url = "https://x\"a\\b" / api_token = "q'q\"b\\s\ntwo" — byte-correct.
• Local full gates re-passed (14/14 entrypoint tests, build, clippy, fmt, codegen).
• CI on 2bb0f57: test: SUCCESS.

This is the second environment-specific shell defect this PR has caught (after the TOML literal-string bug on merged main) — both are now regression-locked by the 14-test suite running the real script through the production parser.

[shivros]

Auto-Merge Gate — eligible (0.94)

Rationale: COD-470's focused two-file fix replaces invalid TOML literal-string handling with escaped TOML basic strings for both Iris fields. The regression suite executes the real entrypoint through Rite's production parser, including quote, backslash, CRLF/newline, control-byte, DEL, and no-token-leak cases.

Checks observed: GitHub CI test succeeded; local gate re-ran cargo test -p rite-server --test docker_entrypoint (14 passed) and cargo fmt --all -- --check.

Scope limits: No deployment/cutover, credentials, or destructive behavior is included. Linked Linear ticket has no manual-review marker and its acceptance criteria are satisfied by this PR.